GUIDE User Manual

(updated for version 26.0)

Wei-Yin Loh
Department of Statistics
University of WisconsinMadison
June 3, 2017

Contents
1 Warranty disclaimer 4

2 Introduction 5
2.1 Installation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 LATEX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Program operation 10
3.1 Required files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4 Classification 13
4.1 Univariate splits, ordinal predictors: glaucoma data . . . . . . . . . . 13
4.1.1 Input file generation . . . . . . . . . . . . . . . . . . . . . . . 14
4.1.2 Contents of glaucoma.in . . . . . . . . . . . . . . . . . . . . 15
4.1.3 Executing the program . . . . . . . . . . . . . . . . . . . . . . 16
4.1.4 Interpreting the output file . . . . . . . . . . . . . . . . . . . . 19
4.2 Linear splits: glaucoma data . . . . . . . . . . . . . . . . . . . . . . . 26

Based on work partially supported by grants from the U.S. Army Research Office, National
Science Foundation, National Institutes of Health, Bureau of Labor Statistics, and Eli Lilly & Co.
Work on precursors to GUIDE additionally supported by IBM and Pfizer.

1
CONTENTS CONTENTS

4.3 Univariate splits, categorical predictors: peptide data . . . . . . . . . 30

4.3.1 Input file generation . . . . . . . . . . . . . . . . . . . . . . . 30
4.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.4 Unbalanced classes and equal priors: hepatitis data . . . . . . . . . . 35
4.5 Unequal misclassification costs: hepatitis data . . . . . . . . . . . . . 39
4.6 More than 2 classes: dermatology . . . . . . . . . . . . . . . . . . . . 40
4.6.1 Default option . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.6.2 Nearest-neighbor option . . . . . . . . . . . . . . . . . . . . . 49
4.6.3 Kernel density option . . . . . . . . . . . . . . . . . . . . . . . 58
4.7 More than 2 classes: heart disease . . . . . . . . . . . . . . . . . . . . 69
4.7.1 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.7.2 GUIDE results . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.7.3 RPART model . . . . . . . . . . . . . . . . . . . . . . . . . . 92

5 Regression 92
5.1 Least squares constant: birthwt data . . . . . . . . . . . . . . . . . . 93
5.1.1 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.1.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.1.3 Contents of cons.var . . . . . . . . . . . . . . . . . . . . . . 108
5.2 Least squares simple linear: birthwt data . . . . . . . . . . . . . . . . 109
5.2.1 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.2.3 Contents of lin.var . . . . . . . . . . . . . . . . . . . . . . . 117
5.2.4 Contents of lin.reg . . . . . . . . . . . . . . . . . . . . . . . 117
5.3 Multiple linear: birthwt data . . . . . . . . . . . . . . . . . . . . . . . 117
5.3.1 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.3.3 Contents of mul.var . . . . . . . . . . . . . . . . . . . . . . . 124
5.3.4 Contents of mul.reg . . . . . . . . . . . . . . . . . . . . . . . 125
5.4 Stepwise linear: birthwt data . . . . . . . . . . . . . . . . . . . . . . 125
5.4.1 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.4.2 Contents of step.reg . . . . . . . . . . . . . . . . . . . . . . 128
5.5 Best ANCOVA: birthwt data . . . . . . . . . . . . . . . . . . . . . . 128
5.5.1 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.5.3 Contents of ancova.reg . . . . . . . . . . . . . . . . . . . . . 135
5.6 Quantile regression: birthwt data . . . . . . . . . . . . . . . . . . . . 135
5.6.1 Piecewise constant: 1 quantile . . . . . . . . . . . . . . . . . . 135

Wei-Yin Loh 2 GUIDE manual

CONTENTS CONTENTS

5.6.2 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 135

5.6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.6.4 Piecewise constant: 2 quantiles . . . . . . . . . . . . . . . . . 143
5.6.5 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 143
5.6.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
5.6.7 Piecewise simple linear . . . . . . . . . . . . . . . . . . . . . . 153
5.6.8 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 153
5.6.9 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
5.6.10 Piecewise multiple linear . . . . . . . . . . . . . . . . . . . . . 160
5.6.11 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 160
5.6.12 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5.7 Least median of squares: birthwt data . . . . . . . . . . . . . . . . . 164
5.7.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
5.8 Poisson regression with offset: lung cancer data . . . . . . . . . . . . 172
5.8.1 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 174
5.8.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.9 Censored response: heart attack data . . . . . . . . . . . . . . . . . . 179
5.9.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
5.10 Multi-response: public health data . . . . . . . . . . . . . . . . . . . 188
5.10.1 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 189
5.10.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
5.11 Longitudinal response with varying time: wage data . . . . . . . . . . 195
5.11.1 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 197
5.11.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
5.12 Subgroup identification: breast cancer . . . . . . . . . . . . . . . . . 206
5.12.1 Without linear prognostic control . . . . . . . . . . . . . . . . 206
5.12.2 With linear prognostic control . . . . . . . . . . . . . . . . . . 213

6 Importance scoring 219

6.1 Classification: glaucoma data . . . . . . . . . . . . . . . . . . . . . . 219
6.1.1 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . 219
6.1.2 Contents of imp.out . . . . . . . . . . . . . . . . . . . . . . . 221
6.2 Regression with censoring: heart attack data . . . . . . . . . . . . . . 223

7 Differential item functioning: GDS data 227

Wei-Yin Loh 3 GUIDE manual

 1 WARRANTY DISCLAIMER

8 Tree ensembles 230

8.1 GUIDE forest: hepatitis data . . . . . . . . . . . . . . . . . . . . . . 231
8.2 Input file creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
8.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
8.4 Bagged GUIDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

9 Other features 239

9.1 Pruning with test samples . . . . . . . . . . . . . . . . . . . . . . . . 239
9.2 Prediction of test samples . . . . . . . . . . . . . . . . . . . . . . . . 239
9.3 GUIDE in R and in simulations . . . . . . . . . . . . . . . . . . . . . 239
9.4 Generation of powers and products . . . . . . . . . . . . . . . . . . . 240
9.5 Data formatting functions . . . . . . . . . . . . . . . . . . . . . . . . 241

1 Warranty disclaimer
Redistribution and use in binary forms, with or without modification, are permitted
provided that the following condition is met:
Redistributions in binary form must reproduce the above copyright notice, this
condition and the following disclaimer in the documentation and/or other materials
provided with the distribution.
THIS SOFTWARE IS PROVIDED BY WEI-YIN LOH AS IS AND ANY EX-
PRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL WEI-YIN
LOH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EX-
EMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIM-
ITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
The views and conclusions contained in the software and documentation are those
of the author and should not be interpreted as representing official policies, either
expressed or implied, of the University of Wisconsin.

Wei-Yin Loh 4 GUIDE manual

 2 INTRODUCTION

2 Introduction
GUIDE stands for Generalized, Unbiased, Interaction Detection and Estimation. It
is the only classification and regression tree algorithm with all these features:

1. Unbiased variable selection.

2. Kernel and nearest-neighbor node models for classification trees.

3. Weighted least squares, least median of squares, quantile, Poisson, and relative
risk (proportional hazards) regression models.

4. Univariate, multivariate, and longitudinal response variables.

5. Pairwise interaction detection at each node.

6. Linear splits on two variables at a time for classification trees.

7. Categorical variables for splitting only, or for both splitting and fitting (via 0-1
dummy variables), in regression tree models.

8. Ranking and scoring of predictor variables.

9. Tree ensembles (bagging and forests).

Tables 1 and 2 compare the features of GUIDE with CRUISE (Kim and Loh, 2001,
2003), QUEST (Loh and Shih, 1997), C4.5 (Quinlan, 1993), RPART1 , and M5
(Quinlan, 1992; Witten and Frank, 2000).
The GUIDE algorithm is documented in Loh (2002) for regression trees and
Loh (2009) for classification trees. Loh (2008a), Loh (2011) and Loh (2014) review
the subject. Advanced features of the algorithm are reported in Chaudhuri and Loh
(2002), Loh (2006b), Kim et al. (2007), Loh et al. (2007), and Loh (2008b). For a list
of third-party applications of GUIDE, CRUISE, QUEST, and the logistic regression
tree algorithm LOTUS (Chan and Loh, 2004; Loh, 2006a), see
http://www.stat.wisc.edu/~ loh/apps.html
This manual illustrates the use of the program and interpretation of the output.
1
RPART is an implementation of CART (Breiman et al., 1984) in R. CART is a registered
trademark of California Statistical Software, Inc.

Wei-Yin Loh 5 GUIDE manual

 2 INTRODUCTION

Table 1: Comparison of GUIDE, QUEST, CRUISE, CART, and C4.5 classification

tree algorithms. Node models: S = simple, K = kernel, L = linear discriminant, N =
nearest-neighbor.
GUIDE QUEST CRUISE CART C4.5
Unbiased splits Yes Yes Yes No No
Splits per node 2 2 2 2 2
Interaction Yes No Yes No No
detection
Importance Yes No No Yes No
ranking
Class priors Yes Yes Yes Yes No
Misclassification Yes Yes Yes Yes No
costs
Linear splits Yes Yes Yes Yes No
Categorical Subsets Subsets Subsets Subsets Atoms
splits
Node models S, K, N S S, L S S
Missing values Special Imputation Surrogate Surrogate Weights
Tree diagrams Text and LATEX Proprietary Text
Bagging Yes No No No No
Forests Yes No No No No

Wei-Yin Loh 6 GUIDE manual

 2 INTRODUCTION

Table 2: Comparison of GUIDE, CART and M5 regression tree algorithms

GUIDE CART M5
Unbiased splits Yes No No
Pairwise interac- Yes No No
tion detection
Importance scores Yes Yes No
Loss functions Weighted least squares, least Least squares, Least squares
median of squares, quantile, least absolute only
Poisson, proportional hazards deviations
Survival, longitu- Yes, yes, yes No, no, no No, no, no
dinal and multi-
response data
Node models Constant, multiple, stepwise Constant only Constant and
linear, polynomial, ANCOVA stepwise
Linear models Multiple or stepwise (forward N/A Stepwise
and forward-backward)
Variable roles Split only, fit only, both, nei- Split only Split and fit
ther, weight, censored, offset
Categorical vari- Subsets of categorical values Subsets 0-1 variables
able splits
Tree selection Pruning or stopping rules Pruning only Pruning only
Tree diagrams Text and LATEX Proprietary PostScript
Operation modes Interactive and batch Interactive Interactive
and batch
Case weights Yes Yes No
Transformations Powers and products No No
Missing values in Missing values treated as a Surrogate Imputation
split variables special category splits
Missing values in Choice of separate constant N/A Imputation
linear predictors models or mean imputation
Bagging & forests Yes & yes No & no No & no
Subgroup identifi- Yes No No
cation
Data conversions ARFF, C4.5, Minitab, R, No No
SAS, Statistica, Systat, CSV

Wei-Yin Loh 7 GUIDE manual

2.1 Installation 2 INTRODUCTION

2.1 Installation
GUIDE is available free from www.stat.wisc.edu/~ loh/guide.html in the form of
compiled 32- and 64-bit executables for Linux, Mac OS X, and Windows on Intel
and compatible processors. Data and description files used in this manual are in the
zip file www.stat.wisc.edu/~ loh/treeprogs/guide/datafiles.zip.

Linux: There are three 64-bit executables to choose from: Intel and NAG (for Red
Hat 6.8), and Gfortran (for Ubuntu 16.0). Make the unzipped file executable
by issuing this command in a Terminal application in the folder where the file
is located: chmod a+x guide.

Mac OS X: There are three executables to choose from. Make the unzipped file
executable by issuing this command in a Terminal application in the folder
where the file is located: chmod a+x guide

NAG. This version may be the fastest. It requires no additional software

besides the guide.gz file.
Absoft. This version requires no additional software besides the guide.gz file
too.
Gfortran. This version requires Xcode and gfortran to be installed. To
ensure that the gfortran libraries are placed in the right place, follow
these steps:
1. Install Xcode from https://developer.apple.com/xcode/downloads/.
2. Go to http://hpc.sourceforge.net and download file gcc-6.2-bin.tar.gz
to your Downloads folder. The direct link to the file is
http://prdownloads.sourceforge.net/hpc/gcc-6.2-bin.tar.gz?download
3. Open a Terminal window and type (or copy and paste):
(a) cd ~/Downloads
(b) gunzip gcc-6.2-bin.tar.gz
(c) sudo tar -xvf gcc-6.2-bin.tar -C /

Windows: There are four executables to choose from: Intel (64 or 32 bit), Absoft
(64 bit) and Gfortran (64 bit). The 32-bit executable may run a bit faster
but the 64-bit versions can handle larger arrays. Download the 32 or 64-bit
executable guide.zip and unzip it (right-click on file icon and select Extract
all). The resulting file guide.exe may be placed in one of three places:

Wei-Yin Loh 8 GUIDE manual

2.2 LATEX 2 INTRODUCTION

1. top level of your C: drive (where it can be invoked by typing C:\guide in

a terminal windowsee Section 3.1),
2. a folder that contains your data files, or
3. a folder on your search path.

2.2 LATEX
GUIDE uses the public-domain software LATEX (http://www.ctan.org) to produce
tree diagrams. The specific locations are:

Linux: TeX Live http://www.tug.org/texlive/

Mac: MacTeX http://tug.org/mactex/

Windows: proTeXt http://www.tug.org/protext/

After LATEX is installed, a pdf file of a LATEX file, called diagram.tex say, produced
by GUIDE can be obtained by typing these three commands in a terminal window:

1. latex diagram

2. dvips diagram

3. ps2pdf diagram.ps

The first command produces a file called diagram.dvi which the second com-
mand uses to create a postscript file called diagram.ps. The latter can be viewed
and printed if a postscript viewer (such as Preview for the Mac) is installed. If
no postscript viewer is available, the last command can be used to convert the
postscript file into a pdf file, which can be viewed and printed with Adobe Reader.
The file diagram.tex can be edited to change colors, node sizes, etc. See, e.g.,
http://tug.org/PSTricks/main.cgi/.
Windows users: Convert the postscript figure to Enhanced-format Meta File
(emf) format for use in Windows applications such as Word or PowerPoint. There
are many conversion programs are available on the web, such as Graphic Converter
(http://www.graphic-converter.net/) and pstoedit (http://www.pstoedit.net/).

Wei-Yin Loh 9 GUIDE manual

 3 PROGRAM OPERATION

3 Program operation
3.1 Required files
The GUIDE program requires two text files for input.
Data file: This file contains the training sample. Each file record consists of ob-
servations on the response (i.e., dependent) variable, the predictor (i.e., X or
independent) variables, and optional weight and time variables. Entries in each
record are comma, space, or tab delimited (multiple spaces are treated as one
space, but not for commas). A record can occupy more than one line in the
file, but each record must begin on a new line.
Values of categorical variables can contain any ascii character except single
and double quotation marks, which are used to enclose values that contain
spaces and commas. Values can be up to 60 characters long. Class labels are
truncated to 10 characters in tabular displays.
A common problem among first-time users is getting the data file in proper
shape. If the data are in a spreadsheet and there are no empty cells, export
them to a MS-DOS Comma Separated (csv) file (the MS-DOS CSV format
takes care of carriage return and line feed characters properly). If there are
empty cells, a good solution is to read the spreadsheet into R (using read.csv
with proper specification of the na.strings argument), verify that the data are
correctly read, and then export them to a text file using either write.table
or write.csv.

Description file: This provides information about the name and location of the
data file, names and column positions of the variables, and their roles in the
analysis. Different models may be fitted by changing the roles of the variables.
We demonstrate with the text files glaucoma.rdata and glaucoma.dsc from
www.stat.wisc.edu/~ loh/treeprogs/guide/datafiles.zip or from the R
package ipred (Peters and Hothorn, 2015)). The data give the values of 66
variables obtained from a laser scan image of the optic nerve for 85 normal
people and 85 people with glaucoma. The response variable is Class (normal
or glaucoma). The top and bottom lines of the file glaucoma.dsc are:

glaucoma.rdata
NA
2
1 ag n

Wei-Yin Loh 10 GUIDE manual

3.1 Required files 3 PROGRAM OPERATION

2 at n
3 as n
4 an n
5 ai n
:
63 tension n
64 clv n
65 cs n
66 lora n
67 Class d

The 1st line gives the name of the data file. If the latter is not in the current
folder, gives its full path (e.g., "c:\data\glaucoma.rdata") surrounded by
quotes (because it contains backslashes). The 2nd line gives the missing value
code, which can be up to 80 characters long. If it contains non-alphanumeric
characters, it too must be surrounded by quotation marks. A missing value
code must appear in the second line of the file even if there are no missing
values in the data (in which case any character string not present among the
data values can be used). The 3rd line gives the line number of the first data
record in the data file. Because glaucoma.rdata has the variable names in the
first row, a 2 is placed on the third line of glaucoma.dsc. Blank lines in
the data and description files are ignored. The position, name and role of each
variable comes next (in that order), with one line for each variable.
Variable names must begin with an alphabet and be not more than 60 charac-
ters long. If a name contains non-alphanumeric characters, it must be enclosed
in matching single or double quotes. Spaces and the four characters #, %, {,
and } are replaced by dots (periods) if they appear in a name. Variable names
are truncated to 10 characters in tabular output. Leading and trailing spaces
are dropped.
The following roles for the variables are permitted. Lower and upper case
letters are accepted.
b Categorical variable that is used both for splitting and for node modeling in
regression. It is transformed to 0-1 dummy variables for node modeling.
It is converted to c type for classification.
c Categorical variable used for splitting only.
d Dependent variable. Except for multi-response data (see Sec. 5.10), there
can only be one such variable. In the case of relative risk models, this

Wei-Yin Loh 11 GUIDE manual

3.1 Required files 3 PROGRAM OPERATION

is the death indicator. The variable can take character string values for
classification.
f Numerical variable used only for fitting the linear models in the nodes of the
tree. It is not used for splitting the nodes and is disallowed in classification.
n Numerical variable used both for splitting the nodes and for fitting the node
models. It is converted to type s in classification.
r Categorical treatment (Rx) variable used only for fitting the linear models
in the nodes of the tree. It is not used for splitting the nodes. If this
variable is present, all n variables are automatically changed to s.
s Numerical-valued variable only used for splitting the nodes. It is not used as
a regressor in the linear models. This role is suitable for ordinal categorical
variables if they are given numerical values that reflect the orderings.
t Survival time (for proportional hazards models) or observation time (for
longitudinal models) variable.
w Weight variable for weighted least squares regression or for excluding ob-
servations in the training sample from tree construction. See section 9.2
for the latter. Except for longitudinal models, a record with a missing
value in a d, t, or z-variable is automatically assigned zero weight.
x Excluded variable. This allows models to be fitted to different subsets of
the variables without reformatting the data file.
z Offset variable used only in Poisson regression.

GUIDE runs within a terminal window of the computer operating system.

Do not double-click its icon!

Linux. Any terminal program will do.

Mac OS X. The program is called Terminal; it is in the Applications Folder.

Windows. The terminal program is started from the Start button by choosing
All Programs Accessories Command Prompt

After the terminal window is opened, change to the folder where the data and pro-
gram files are stored. For Windows users who do not know how to do this, read
http://www.digitalcitizen.life/command-prompt-how-use-basic-commands.

Wei-Yin Loh 12 GUIDE manual

3.2 Input file creation 4 CLASSIFICATION

3.2 Input file creation

GUIDE is started by typing its (lowercase) name in a terminal. The preferred way
is to create an input file (option 1 below) for subsequent execution. The input file
may be edited if you wish to change some input parameters later. In the following,
the sign (>) is the terminal prompt (not to be typed!).
> guide
GUIDE Classification and Regression Trees and Forests
Version 26.0 (Build date: June 1, 2017)
Compiled with GFortran 6.2.0 on Mac OS X Sierra 10.12.5
Copyright (c) 1997-2017 Wei-Yin Loh. All rights reserved.
This software is based upon work supported by the U.S. Army Research Office,
the National Science Foundation and the National Institutes of Health.

Choose one of the following options:

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice:

The meanings of these options are:

0. Print the warranty disclaimer.

1. Create an input file for model fitting or importance scoring (recommended).

2. Convert the data file into a format suitable for importation into database, spread-
sheet, or statistics software. See Table 2 for the statistical packages supported.
Section 9.5 has an example.

4 Classification
4.1 Univariate splits, ordinal predictors: glaucoma data
We first show how to generate an input file to produce a classification tree from
the data in the file glaucoma.rdata, using the default options. Whenever you are
prompted for a selection, there is usually range of permissible values given within
square brackets and a default choice (indicated by the symbol <cr>=). The default
may be selected by pressing the ENTER or RETURN key. Annotations are printed in
blue italics in this manual.

Wei-Yin Loh 13 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

4.1.1 Input file generation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: glaucoma.in
This file will store your answers to the prompts.
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Press the ENTER or RETURN key to accept the default selection.
Name of batch output file: glaucoma.out
This file will contain the results when you apply the input file to GUIDE later.
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Option 2 is for bagging and random forest-type methods.
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):
The default option will produce a traditional classification tree.
Choose option 2 for more advanced features.
Input name of data description file (max 100 characters);
enclose with matching quotes if it has spaces: glaucoma.dsc
Reading data description file ...
Training sample file: glaucoma.rdata
The name of the data set is read from the description file.
Some information about the data are printed in the next few lines.
Missing value code: NA
Records in data file start on line 2
Warning: N variables changed to S
This warning is due to N variables being always used as S in classification.
Dependent variable is Class
Reading data file ...
Number of records in data file: 170
Length of longest data entry: 8
Checking for missing values ...
Total number of cases: 170
Number of classes = 2
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values
Finished checking data
Creating missing value indicators
Rereading data
Class #Cases Proportion
glaucoma 85 0.50000000

Wei-Yin Loh 14 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

normal 85 0.50000000
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
170 0 17 0 0 0 66 0 0
No. cases used for training: 170
No. cases excluded due to 0 weight or missing D: 0
Finished reading data file
Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3 ([1:3], <cr>=1):
See other parts of manual for examples of equal and specified priors.
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Choose option 2 if you do not want LaTeX code.
Input file name to store LaTeX code (use .tex as suffix): glaucoma.tex
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: glaucoma.fit
This file will contain the node number and predicted class for each observation.
Input file is created!
Run GUIDE with the command: guide < glaucoma.in

4.1.2 Contents of glaucoma.in

Here are the contents of the input file:
GUIDE (do not edit this file unless you know what you are doing)
26.0 (version of GUIDE that generated this file)
1 (1=model fitting, 2=importance or DIF scoring, 3=data conversion)
"glaucoma.out" (name of output file)
1 (1=one tree, 2=ensemble)
1 (1=classification, 2=regression, 3=propensity score grouping)
1 (1=simple model, 2=nearest-neighbor, 3=kernel)
1 (0=linear 1st, 1=univariate 1st, 2=skip linear, 3=skip linear and interaction)
1 (1=prune by CV, 2=by test sample, 3=no pruning)
"glaucoma.dsc" (name of data description file)
10 (number of cross-validations)
1 (1=mean-based CV tree, 2=median-based CV tree)
0.500 (SE number for pruning)
1 (1=estimated priors, 2=equal priors, 3=other priors)
1 (1=unit misclassification costs, 2=other)
2 (1=split point from quantiles, 2=use exhaustive search)
1 (1=default max. number of split levels, 2=specify no. in next line)
1 (1=default min. node size, 2=specify min. value in next line)
1 (1=write latex, 2=skip latex)
"glaucoma.tex" (latex file name)

Wei-Yin Loh 15 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

1 (1=vertical tree, 2=sideways tree)

1 (1=include node numbers, 2=exclude)
1 (1=number all nodes, 2=only terminal nodes)
1 (1=color terminal nodes, 2=no colors)
1 (0=#errors, 1=class sizes in nodes, 2=nothing)
1 (1=no storage, 2=store fit and split variables, 3=store split variables and values)
2 (1=do not save individual fitted values and node IDs, 2=save in a file)
"glaucoma.fit" (file name for fitted values and node IDs)
1 (1=do not write R function, 2=write R function)

GUIDE reads only the first item in each line; the rest of the line is a comment for
human consumption. It is generally not advisable for the user to edit this file because
each question depends on the answers given to previous questions.

4.1.3 Executing the program

After the input file is generated, GUIDE can be executed by typing the command
guide < glaucoma.in at the screen prompt:
> guide < glaucoma.in

This produces the following output to the screen. The alternative command guide
< glaucoma.in > log.txt sends the screen output to the file log.txt.
GUIDE Classification and Regression Trees and Forests
Version 26.0 (Build date: June 1, 2017)
Compiled with GFortran 6.2.0 on Mac OS X Sierra 10.12.5
Copyright (c) 1997-2017 Wei-Yin Loh. All rights reserved.
This software is based upon work supported by the U.S. Army Research Office,
the National Science Foundation and the National Institutes of Health.

Choose one of the following options:

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: Batch run with input file
Input 1 for model fitting, 2 for importance or DIF scoring, 3 for data conversion: 1
Output file is glaucoma.out

Input 1 for single tree, 2 for ensemble of trees: 1

Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice: 1
Input 1 for simple, 2 for nearest-neighbor, 3 for kernel method: 1

Wei-Yin Loh 16 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

Input 0 for linear, interaction and univariate splits (in this order),
1 for univariate, linear and interaction splits (in this order),
2 to skip linear splits,
3 to skip linear and interaction splits: 1
Input 1 to prune by CV, 2 by test sample, 3 for no pruning: 1

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: glaucoma.dsc
Reading data description file ...
Training sample file: glaucoma.rdata
Missing value code: NA
Records in data file start on line 2
Warning: N variables changed to S
Dependent variable is Class
Reading data file ...
Number of records in data file: 170
Length of longest data entry: 8
Checking for missing values ...
Total number of cases: 170
Number of classes: 2
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values
Data checks complete
Creating missing value indicators
Rereading data
Class #Cases Proportion
glaucoma 85 0.50000000
normal 85 0.50000000
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
170 0 17 0 0 0 66 0 0
No. cases used for training: 170
No. cases excluded due to 0 weight or missing D: 0
Finished reading data file
Univariate split highest priority
Interaction and linear splits 2nd and 3rd priorities
Input number of cross-validations: 10
Selected tree is based on mean of CV estimates
Input number of SEs for pruning: 0.50000000000000000
Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3: 1
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2: 1
Choose a split point selection method for numerical variables:
Choose 1 to use faster method based on sample quantiles

Wei-Yin Loh 17 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

Choose 2 to use exhaustive search

Input 1 or 2: 2
Max. number of split levels: 10
Input 1 for default min. node size,
2 to specify min. value: 1
Input 1 for LaTeX tree code, 2 to skip it: 1
Input file name to store LaTeX code: glaucoma.tex
Warning: LaTeX file is overwritten
Input 1 to include node numbers, 2 to omit them: 1
Input 1 to number all nodes, 2 to number leaves only: 1
Input 0 for #errors, 1 for class sizes in nodes, 2 for nothing: 1
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values
Input your choice: 1
Input 2 to save fitted values and node IDs; 1 otherwise: 2
File name is glaucoma.fit
Warning: file is overwritten
Input 2 to write R function for predicting new cases, 1 otherwise: 1
Constructing main tree ...
Number of terminal nodes of largest tree: 8
Performing cross-validation:
Finished cross-validation iteration 1
Finished cross-validation iteration 2
Finished cross-validation iteration 3
Finished cross-validation iteration 4
Finished cross-validation iteration 5
Finished cross-validation iteration 6
Finished cross-validation iteration 7
Finished cross-validation iteration 8
Finished cross-validation iteration 9
Finished cross-validation iteration 10

Pruning main tree. Please wait.

Results of subtree sequence
Trees based on mean with naive SE are marked with * and **
Tree based on mean with bootstrap SE is marked with --
Trees based on median with finite bootstrap SE are marked with + and ++
Subtree #Terminal nodes
0 8
1** 5
2 3
3 2
4 1
* tree, ** tree, + tree, and ++ tree all the same

Wei-Yin Loh 18 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

Results are stored in glaucoma.out

Observed and fitted values are stored in glaucoma.fit
LaTeX code for tree is in glaucoma.tex

The final pruned tree is marked with two asterisks (**); it has 4 terminal nodes.

4.1.4 Interpreting the output file

Following is an annotated copy of the contents of the output file.
Classification tree
Pruning by cross-validation
Data description file: glaucoma.dsc
Training sample file: glaucoma.rdata
Missing value code: NA
Records in data file start on line 2
This says that the first record begins on line 2 of the data file.
Warning: N variables changed to S
This warning is triggered if classification is chosen and there are predictor
variables designated as N.
Dependent variable is Class
Number of records in data file: 170
Length of longest data entry: 8
Class proportions of dependent variable Class:
Number of classes: 2
Class #Cases Proportion
glaucoma 85 0.50000000
normal 85 0.50000000
This gives the number of observations in each class.
Summary information (without x variables)
d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 ag s 1.1220E+00 4.6110E+00
2 at s 1.7600E-01 9.2400E-01
3 as s 3.0800E-01 1.1730E+00
4 an s 3.4500E-01 1.5640E+00
5 ai s 2.9400E-01 1.1250E+00
6 eag s 4.1500E-01 3.9480E+00
7 eat s 1.3700E-01 8.4800E-01
8 eas s 4.3000E-02 1.0610E+00
9 ean s 8.0000E-03 1.2660E+00
10 eai s 9.8000E-02 9.6100E-01
11 abrg s 3.0000E-03 3.8940E+00

Wei-Yin Loh 19 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

12 abrt s 3.0000E-03 8.2700E-01

13 abrs s 0.0000E+00 9.0100E-01
14 abrn s 0.0000E+00 1.2680E+00
15 abri s 0.0000E+00 9.1500E-01
16 hic s -1.8900E-01 8.8700E-01
17 mhcg s -1.4700E-01 3.2200E-01
18 mhct s -4.7000E-02 4.7700E-01
19 mhcs s -1.7200E-01 2.9300E-01
20 mhcn s -2.1200E-01 3.8500E-01
21 mhci s -1.6100E-01 4.5400E-01
22 phcg s -2.8600E-01 1.4500E-01
23 phct s -1.2100E-01 4.0200E-01
24 phcs s -2.4700E-01 1.6000E-01
25 phcn s -2.8500E-01 2.1700E-01
26 phci s -2.8600E-01 3.7100E-01
27 hvc s 1.1000E-01 7.1500E-01
28 vbsg s 2.0000E-02 2.0770E+00
29 vbst s 7.0000E-03 4.4600E-01
30 vbss s 2.0000E-03 5.5400E-01
31 vbsn s 0.0000E+00 6.9600E-01
32 vbsi s 6.0000E-03 4.9000E-01
33 vasg s 5.0000E-03 7.5100E-01
34 vast s 0.0000E+00 1.5000E-02
35 vass s 1.0000E-03 2.3900E-01
36 vasn s 1.0000E-03 3.9700E-01
37 vasi s 1.0000E-03 1.0500E-01
38 vbrg s 0.0000E+00 1.9890E+00
39 vbrt s 0.0000E+00 3.9900E-01
40 vbrs s 0.0000E+00 5.4400E-01
41 vbrn s 0.0000E+00 6.7900E-01
42 vbri s 0.0000E+00 4.2800E-01
43 varg s 6.0000E-03 1.3250E+00
44 vart s 1.0000E-03 6.5000E-02
45 vars s 3.0000E-03 3.9700E-01
46 varn s 1.0000E-03 5.9700E-01
47 vari s 0.0000E+00 2.6600E-01
48 mdg s 1.2100E-01 1.2980E+00
49 mdt s 1.1700E-01 1.2150E+00
50 mds s 1.3700E-01 1.3510E+00
51 mdn s 2.3000E-02 1.2600E+00
52 mdi s 1.1600E-01 1.2470E+00
53 tmg s -3.5300E-01 1.9200E-01
54 tmt s -2.5900E-01 3.6600E-01
55 tms s -4.3000E-01 3.5800E-01
56 tmn s -5.1000E-01 2.4500E-01
57 tmi s -4.0500E-01 2.8600E-01

Wei-Yin Loh 20 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

58 mr s 5.9900E-01 1.2190E+00
59 rnf s -1.9000E-02 4.5100E-01
60 mdic s 1.2000E-02 6.6300E-01
61 emd s 4.7000E-02 7.4300E-01
62 mv s 0.0000E+00 1.8300E-01
63 tension s 1.0000E+01 2.5000E+01 4
64 clv s 0.0000E+00 1.4600E+02 12
65 cs s 3.3000E-01 1.9100E+00 1
66 lora s 0.0000E+00 9.2578E+01
67 Class d 2
This shows the type, minimum, maximum and number of missing values of each variable.
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
170 0 17 0 0 0 66 0 0
This shows the number of each type of variable.
No. cases used for training: 170
No. cases excluded due to 0 weight or missing D: 0

Univariate split highest priority

Interaction and linear splits 2nd and 3rd priorities
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Simple node models
Estimated priors
Unit misclassification costs
Split values for N and S variables based on exhaustive search
Max. number of split levels: 10
Min. node sample size: 2
Number of SEs for pruned tree: 5.0000E-01

Size and CV mean cost and SE of subtrees:

Tree #Tnodes Mean Cost SE(Mean) BSE(Mean) Median Cost BSE(Median)
0 8 9.412E-02 2.239E-02 3.551E-02 5.882E-02 3.213E-02
1** 5 8.824E-02 2.175E-02 3.674E-02 5.882E-02 3.824E-02
2 3 1.235E-01 2.524E-02 2.543E-02 1.176E-01 2.792E-02
3 2 1.471E-01 2.716E-02 2.436E-02 1.471E-01 3.576E-02
4 1 5.000E-01 3.835E-02 9.213E-03 5.000E-01 2.508E-02

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree and ++ tree are the same
The tree with the smallest mean CV cost is marked with an asterisk.
The selected tree is marked with two asterisks; it is the smallest one

Wei-Yin Loh 21 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

having mean CV cost within the specified standard error (SE) bounds.
The mean CV costs and SEs are given in the 3rd and 4th columns.
The other columns are bootstrap estimates used for experimental purposes.

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Node cost is node misclassification cost divided by number of training cases

Node Total Train Predicted Node Split Interacting
label cases cases class cost variables variable
1 170 170 glaucoma 5.000E-01 lora
2 73 73 normal 9.589E-02 clv
4T 62 62 normal 0.000E+00 -
5 11 11 glaucoma 3.636E-01 lora
10T 4 4 normal 0.000E+00 -
11T 7 7 glaucoma 0.000E+00 -
3 97 97 glaucoma 1.959E-01 clv
6T 15 15 normal 6.667E-02 -
7T 82 82 glaucoma 6.098E-02 vass :clv
This shows the tree structure in tabular form. A node with label k has its left
and right child nodes are labeled 2k and 2k+1, respectively. Terminal nodes are
indicated with the symbol T. The notation :tmi at node 7 indicates that
the variable clv has an interaction with the split variable vass.

Number of terminal nodes of final tree: 5

Total number of nodes of final tree: 9
Second best split variable (based on curvature test) at root node is clv
This says that lora is the second best variable to split the root node.
Classification tree:
The tree structure is shown next in indented text form.
Node 1: lora <= 56.40073
Node 2: clv <= 8.40000 or NA
Node 4: normal
Node 2: clv > 8.40000
Node 5: lora <= 49.23372
Node 10: normal
Node 5: lora > 49.23372 or NA
Node 11: glaucoma
Node 1: lora > 56.40073 or NA
Node 3: clv <= 2.00000
Node 6: normal
Node 3: clv > 2.00000 or NA
Node 7: glaucoma

***************************************************************

Wei-Yin Loh 22 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

Node compositions and other details are given next.

In the following the predictor node mean is mean of complete cases.

Node 1: Intermediate node

A case goes into Node 2 if lora <= 5.6400730E+01
lora mean = 5.7555E+01
Class Number ClassPrior
glaucoma 85 0.50000
normal 85 0.50000
Number of training cases misclassified = 85
Predicted class is glaucoma
----------------------------
Node 2: Intermediate node
A case goes into Node 4 if clv <= 8.4000000E+00 or NA
clv mean = 5.4861E+00
Class Number ClassPrior
glaucoma 7 0.09589
normal 66 0.90411
Number of training cases misclassified = 7
Predicted class is normal
----------------------------
Node 4: Terminal node
Class Number ClassPrior
glaucoma 0 0.00000
normal 62 1.00000
Number of training cases misclassified = 0
Predicted class is normal
----------------------------
Node 5: Intermediate node
A case goes into Node 10 if lora <= 4.9233715E+01
lora mean = 4.9100E+01
Class Number ClassPrior
glaucoma 7 0.63636
normal 4 0.36364
Number of training cases misclassified = 4
Predicted class is glaucoma
----------------------------
Node 10: Terminal node
Class Number ClassPrior
glaucoma 0 0.00000
normal 4 1.00000
Number of training cases misclassified = 0
Predicted class is normal
----------------------------
Node 11: Terminal node
Class Number ClassPrior

Wei-Yin Loh 23 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

glaucoma 7 1.00000
normal 0 0.00000
Number of training cases misclassified = 0
Predicted class is glaucoma
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if clv <= 2.0000000E+00
clv mean = 3.5821E+01
Class Number ClassPrior
glaucoma 78 0.80412
normal 19 0.19588
Number of training cases misclassified = 19
Predicted class is glaucoma
----------------------------
Node 6: Terminal node
Class Number ClassPrior
glaucoma 1 0.06667
normal 14 0.93333
Number of training cases misclassified = 1
Predicted class is normal
----------------------------
Node 7: Terminal node
Class Number ClassPrior
glaucoma 77 0.93902
normal 5 0.06098
Number of training cases misclassified = 5
Predicted class is glaucoma
----------------------------

Classification matrix for training sample:

Predicted True class
class glaucoma normal
glaucoma 84 5
normal 1 80
Total 85 85

Number of cases used for tree construction: 170

Number misclassified: 6
Resubstitution est. of mean misclassification cost: 0.3529E-01

Observed and fitted values are stored in glaucoma.fit

LaTeX code for tree is in glaucoma.tex

Figure 1 shows the classification tree drawn by LaTeX using the file glaucoma.tex.
The last sentence in its caption gives the second best variable for splitting the root

Wei-Yin Loh 24 GUIDE manual

4.1 Univariate splits, ordinal predictors: glaucoma data 4 CLASSIFICATION

lora
56.40 1

clv clv
8.40 2 2.00 3

0 lora 1 77
62 4 49.23 5 14 6 7 5
normal normal glaucoma
0 7
4 10 11 0
normal glaucoma
150

glaucoma
normal
100
clv

50
0

0 20 40 60 80

lora
Figure 1: GUIDE v.26.0 0.50-SE classification tree for predicting Class using esti-
mated priors and unit misclassification costs. At each split, an observation goes to
the left branch if and only if the condition is satisfied. The symbol stands for
or missing. Predicted classes (based on estimated misclassification cost) printed
below terminal nodes; sample sizes for Class = glaucoma and normal, respectively,
beside nodes. Second best split variable at root node is clv.
Wei-Yin Loh 25 GUIDE manual
4.2 Linear splits: glaucoma data 4 CLASSIFICATION

node. The top lines of the file glaucoma.fit are shown below. Their order corre-
sponds to the order of the observations in the training sample file. The 1st column
(labeled train) indicates whether the observation is used (y) or not used (n) to
fit the model. Since we used the entire data set to fit the model here, all the entries
in the first column are y. The 2nd column gives the terminal node number that the
observation belongs to and the 3rd and 4th columns give its observed and predicted
classes.
train node observed predicted
y 4 "normal" "normal"
y 4 "normal" "normal"
y 4 "normal" "normal"
y 4 "normal" "normal"
y 6 "normal" "normal"

4.2 Linear splits: glaucoma data

This section shows how to make GUIDE use linear splits on two variables at a time.
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: lin.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: lin.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):2
Choosing 2 enables more options.
Input 1 for simple, 2 for nearest-neighbor, 3 for kernel method
([1:3], <cr>=1):
Options 2 and 3 yield nearest-neighbor and kernel discriminant node models.
Input 0 for linear, interaction and univariate splits (in this order),
1 for univariate, linear and interaction splits (in this order),
2 to skip linear splits,
3 to skip linear and interaction splits:
Input your choice ([0:3], <cr>=1):0
Option 1 is the default.
Input 1 to prune by CV, 2 by test sample, 3 for no pruning ([1:3], <cr>=1):

Input name of data description file (max 100 characters);

Wei-Yin Loh 26 GUIDE manual

4.2 Linear splits: glaucoma data 4 CLASSIFICATION

enclose with matching quotes if it has spaces: glaucoma.dsc

Reading data description file ...
Training sample file: glaucoma.rdata
Missing value code: NA
Records in data file start on line 2
Warning: N variables changed to S
Dependent variable is Class
Reading data file ...
Number of records in data file: 170
Length of longest data entry: 8
Checking for missing values ...
Total number of cases: 170
Number of classes: 2
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values
Finished checking data
Creating missing value indicators
Rereading data
Class #Cases Proportion
glaucoma 85 0.50000000
normal 85 0.50000000
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
170 0 17 0 0 0 66 0 0
No. cases used for training: 170
No. cases excluded due to 0 weight or missing D: 0
Finished reading data file
Default number of cross-validations: 10
Input 1 to accept the default, 2 to change it ([1:2], <cr>=1):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50):
Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3 ([1:3], <cr>=1):
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1):
Choose a split point selection method for numerical variables:
Choose 1 to use faster method based on sample quantiles
Choose 2 to use exhaustive search
Input 1 or 2 ([1:2], <cr>=2):
Default max number of split levels: 10
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 10
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):

Wei-Yin Loh 27 GUIDE manual

4.2 Linear splits: glaucoma data 4 CLASSIFICATION

Input file name to store LaTeX code (use .tex as suffix): lin.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1):
Choosing 2 will give a tree with no node labels.
Input 1 to number all nodes, 2 to number leaves only ([1:2], <cr>=1):
Input 1 to color terminal nodes, 2 otherwise ([1:2], <cr>=1):
Choose amount of detail in nodes of LaTeX tree diagram
Input 0 for #errors, 1 for class sizes, 2 for nothing ([0:2], <cr>=1):
Choose 2 if a large tree is expected.
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split variables and their values
Input your choice ([1:2], <cr>=1): 2
Choose 2 to output the info to another file for further processing.
Input file name: linvar.txt
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: lin.fit
Input 2 to save terminal node IDs for importance scoring; 1 otherwise ([1:2], <cr>=1):
Input 2 to write R function for predicting new cases, 1 otherwise ([1:2], <cr>=1):2
Input file name: linpred.r
Input file is created!
Run GUIDE with the command: guide < lin.in

Running GUIDE with the input file yields the following results. The LATEX tree
diagram and partitions are shown in Figure 2.
Node 1: 4.1110165E-01 * clv + lora <= 5.9402920E+01
Node 2: normal
Node 1: 4.1110165E-01 * clv + lora > 5.9402920E+01 or NA
Node 3: glaucoma

Contents of linvar.txt: This file gives information about the splits:

1 1 lora clv 2 0.4111016476E+00 0.5940292030E+02
2 t mdn clv "normal"
3 t cs ean "glaucoma"

Each row refers to a node. The 1st column gives the node number. The 2nd column
contains the letter 1, n, s, c, or t, indicating a split on two variables, a n variable, a s
variable, a c variable, or a terminal node, respectively. The 3rd and 4th columns give
the names of the 2 variables in a bivariate split or the names of the split variable and
the interacting variable in a univariate split. If a node cannot be split, the words
NONE are printed. If a node is terminal, the predicted class is printed in the 5th
column. Otherwise, if it is a non-terminal node, the 5th column gives the number of
values to follow. In the above example, the 2 in the 5th column of each non-terminal
node indicates that it is followed by two parameter values defining the linear split. If

Wei-Yin Loh 28 GUIDE manual

4.2 Linear splits: glaucoma data 4 CLASSIFICATION

lora
clv 1

2 83
71 2 3 14
normal glaucoma
150
100

glaucoma
clv

normal
50
0

0 20 40 60 80

lora
Figure 2: GUIDE v.26.0 0.50-SE classification tree for predicting Class using linear
split priority, estimated priors and unit misclassification costs. At each split, an
observation goes to the left branch if and only if the condition is satisfied. Predicted
classes (based on estimated misclassification cost) printed below terminal nodes;
sample sizes for Class = glaucoma and normal, respectively, beside nodes.

Wei-Yin Loh 29 GUIDE manual

4.3 Univariate splits, categorical predictors: peptide data 4 CLASSIFICATION

the split is on a categorical variable, the 5th column gives the number of categorical
values defining the split and the 6th and subsequent columns give their values.

Contents of linpred.r: This file contains the following R function for predicting
future observations:
predicted <- function(){
if(!is.na(lora) & !is.na(clv) & 0.41110164757186696*clv + lora <= 59.402920297324165){
nodeid <- 2
predict <- "normal"
} else {
nodeid <- 3
predict <- "glaucoma"
}
return(c(nodeid,predict))
}

4.3 Univariate splits, categorical predictors: peptide data

GUIDE can be used with categorical (i.e., nominal) predictor variables as well. We
show this with a data set on peptide binding analyzed by Segal (1988) who used
CART. The data consist of observations on 310 peptides, 181 of which bind to a
Class I MHC molecule and 129 do not. The data are in the file peptide.rdata.
Column 1 contains the peptide ID and column 2 its binding status (bind). The
remaining 112 columns are predictor variables, all continuous except for the last 8
which are categorical (named pos1pos8), each taking 1820 nominal values. Our
goal here is to build a model to predict bind from these 8 categorical variables.
The GUIDE description is peptide.dsc. Note that the 3rd line of the file is
2, indicating that the data begin on line 2 of peptide.rdata (the first line of the
latter contain the names of the variables). Note also that the continuous variables
are excluded from the model by designating each of them with an x.

4.3.1 Input file generation

We use all the default options to produce a GUIDE input file.
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: peptide.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):

Wei-Yin Loh 30 GUIDE manual

4.3 Univariate splits, categorical predictors: peptide data 4 CLASSIFICATION

Name of batch output file: peptide.out

Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: peptide.dsc
Reading data description file ...
Training sample file: peptide.rdata
Missing value code: NA
Records in data file start on line 2
Dependent variable is bind
Reading data file ...
Number of records in data file: 310
Length of longest data entry: 6
Checking for missing values ...
Total number of cases: 310
Number of classes = 2
Col. no. Categorical variable #levels #missing values
107 pos1 18 0
108 pos2 20 0
109 pos3 20 0
110 pos4 20 0
111 pos5 20 0
112 pos6 20 0
113 pos7 19 0
114 pos8 20 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished checking data
Rereading data
Class #Cases Proportion
0 129 0.41612903
1 181 0.58387097
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
310 0 0 105 0 0 0 0 8
No. cases used for training: 310
Finished reading data file
Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3 ([1:3], <cr>=1):
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):

Wei-Yin Loh 31 GUIDE manual

4.3 Univariate splits, categorical predictors: peptide data 4 CLASSIFICATION

Input file name to store LaTeX code (use .tex as suffix): peptide.tex
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: peptide.fit
Input file is created!
Run GUIDE with the command: guide < peptide.in

4.3.2 Results
Results from the output file peptide.out follow.
Classification tree
Pruning by cross-validation
Data description file: segal.dsc
Training sample file: segal.dat
Missing value code: NA
Records in data file start on line 2
Dependent variable is bind
Number of records in data file: 310
Length of longest data entry: 6
Class proportions of dependent variable bind:
Number of classes: 2
Class #Cases Proportion
0 129 0.41612903
1 181 0.58387097

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
2 bind d 2
107 pos1 c 18
108 pos2 c 20
109 pos3 c 20
110 pos4 c 20
111 pos5 c 20
112 pos6 c 20
113 pos7 c 19
114 pos8 c 20

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
310 0 0 105 0 0 0 0 8
No. cases used for training: 310

Wei-Yin Loh 32 GUIDE manual

4.3 Univariate splits, categorical predictors: peptide data 4 CLASSIFICATION

Univariate split highest priority

Interaction and linear splits 2nd and 3rd priorities
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Simple node models
Estimated priors
Unit misclassification costs
Split values for N and S variables based on exhaustive search
Max. number of split levels: 10
Min. node sample size: 2
Number of SEs for pruned tree: 5.0000E-01

Size and CV mean cost and SE of subtrees:

Tree #Tnodes Mean Cost SE(Mean) BSE(Mean) Median Cost BSE(Median)
1 10 1.097E-01 1.775E-02 2.409E-02 9.677E-02 2.879E-02
2 8 1.097E-01 1.775E-02 2.409E-02 9.677E-02 2.879E-02
3 6 1.097E-01 1.775E-02 2.242E-02 9.677E-02 2.416E-02
4 5 1.097E-01 1.775E-02 2.206E-02 8.065E-02 2.231E-02
5 3 1.194E-01 1.841E-02 2.150E-02 1.129E-01 2.576E-02
6** 2 1.097E-01 1.775E-02 2.286E-02 8.065E-02 2.670E-02
7 1 4.161E-01 2.800E-02 3.207E-03 4.194E-01 1.019E-03

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree and ++ tree are the same

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Node cost is node misclassification cost divided by number of training cases

Node Total Train Predicted Node Split Interacting
label cases cases class cost variables variable
1 310 310 1 4.161E-01 pos5
2T 169 169 1 5.917E-02 pos1
3T 141 141 0 1.560E-01 pos8

Number of terminal nodes of final tree: 2

Total number of nodes of final tree: 3
Second best split variable (based on curvature test) at root node is pos1

Classification tree:
At each categorical variable split, values not in training data go right

Wei-Yin Loh 33 GUIDE manual

4.3 Univariate splits, categorical predictors: peptide data 4 CLASSIFICATION

Node 1: pos5 = "F", "M", "Y"

Node 2: 1
Node 1: pos5 /= "F", "M", "Y"
Node 3: 0

***************************************************************

Node 1: Intermediate node

A case goes into Node 2 if pos5 =
"F", "M", "Y"
pos5 mode = "Y"
Class Number ClassPrior
0 129 0.41613
1 181 0.58387
Number of training cases misclassified = 129
Predicted class is 1
----------------------------
Node 2: Terminal node
Class Number ClassPrior
0 10 0.05917
1 159 0.94083
Number of training cases misclassified = 10
Predicted class is 1
----------------------------
Node 3: Terminal node
Class Number ClassPrior
0 119 0.84397
1 22 0.15603
Number of training cases misclassified = 22
Predicted class is 0
----------------------------

Classification matrix for training sample:

Predicted True class
class 0 1
0 119 22
1 10 159
Total 129 181

Number of cases used for tree construction: 310

Number misclassified: 32
Resubstitution est. of mean misclassification cost: 0.1032

Wei-Yin Loh 34 GUIDE manual

4.4 Unbalanced classes and equal priors: hepatitis data 4 CLASSIFICATION

pos5
in S1

10 119
159 22
1 0
Figure 3: GUIDE v.26.0 0.50-SE classification tree for predicting bind using esti-
mated priors and unit misclassification costs. At each split, an observation goes to
the left branch if and only if the condition is satisfied. For splits on categorical vari-
ables, values not present in the training sample go to the right. Set S1 = {F, M, Y}.
Predicted classes (based on estimated misclassification cost) printed below terminal
nodes; sample sizes for bind = 0 and 1, respectively, beside nodes. Second best split
variable at root node is pos1.

Observed and fitted values are stored in uni.fit

LaTeX code for tree is in uni.tex

The results indicate that the largest tree before pruning has 10 terminal nodes.
The pruned tree (marked by **) has 2 terminal nodes. Its cross-validation estimate
of misclassification cost (or error rate here) is 0.1097. Figure 3 shows the pruned
tree. It splits on pos5, sending values A, C, D, E, G, H, I, K, L, N, P, Q, R,
S, T, V, and W to the left node. The second best variable to split the root node is
pos1.

4.4 Unbalanced classes and equal priors: hepatitis data

If a data set has one dominant class, a classification tree may be null after pruning,
as it may be hard to beat the classifier that predicts every observation to belong to
the dominant class. Nonetheless, it may be of interest to find out which variables
are more predictive and how they affect the dependent variable. One solution is to
use the equal priors option. The resulting model should not be used for prediction.
Instead, by comparing the class proportions in each terminal node against those
at the root node, it can be used to identify the nodes where the dominant class
proportion is much higher or much lower than average (i.e., at the root node).
We use a hepatitis data set to show this. The files are hepdsc.txt and hepdat.txt;
see http://archive.ics.uci.edu/ml/datasets/Hepatitis. The data consist of
observations from 155 individuals, of whom 32 are labeled die and 123 labeled

Wei-Yin Loh 35 GUIDE manual

4.4 Unbalanced classes and equal priors: hepatitis data 4 CLASSIFICATION

live. That is, 79% of the individuals are in the live class. The contents of
hepdsc.txt are:
hepdat.txt
"?"
1
1 CLASS d
2 AGE n
3 SEX c
4 STEROID c
5 ANTIVIRALS c
6 FATIGUE c
7 MALAISE c
8 ANOREXIA c
9 BIGLIVER c
10 FIRMLIVER c
11 SPLEEN c
12 SPIDERS c
13 ASCITES c
14 VARICES c
15 BILIRUBIN n
16 ALKPHOSPHATE n
17 SGOT n
18 ALBUMIN n
19 PROTIME n
20 HISTOLOGY c

Using the default estimated priors yields a null tree with no splits. To obtain a
nonnull tree, we choose equal priors here.
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: hepeq.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: hepeq.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1): 2
Option 2 is needed for equal or specified priors.
Input 1 for simple, 2 for nearest-neighbor, 3 for kernel method ([1:3], <cr>=1):
Input 0 for linear, interaction and univariate splits (in this order),
1 for univariate, linear and interaction splits (in this order),

Wei-Yin Loh 36 GUIDE manual

4.4 Unbalanced classes and equal priors: hepatitis data 4 CLASSIFICATION

2 to skip linear splits,

3 to skip linear and interaction splits:
Input your choice ([0:3], <cr>=1):
Input 1 to prune by CV, 2 by test sample, 3 for no pruning ([1:3], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: hepdsc.txt
Reading data description file ...
Training sample file: hepdat.txt
Missing value code: ?
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is CLASS
Reading data file ...
Number of records in data file: 155
Length of longest data entry: 6
Checking for missing values ...
Total number of cases: 155
Missing values found among categorical variables
Separate categories will be created for missing categorical variables
Number of classes: 2
Col. no. Categorical variable #levels #missing values
3 SEX 2 0
4 STEROID 2 1
5 ANTIVIRALS 2 0
6 FATIGUE 2 1
7 MALAISE 2 1
8 ANOREXIA 2 1
9 BIGLIVER 2 10
10 FIRMLIVER 2 11
11 SPLEEN 2 5
12 SPIDERS 2 5
13 ASCITES 2 5
14 VARICES 2 5
20 HISTOLOGY 2 0
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values
Data checks complete
Creating missing value indicators
Rereading data
Class #Cases Proportion
die 32 0.20645161
live 123 0.79354839
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var

Wei-Yin Loh 37 GUIDE manual

4.4 Unbalanced classes and equal priors: hepatitis data 4 CLASSIFICATION

155 0 72 0 0 0 6 0 13
No. cases used for training: 155
No. cases excluded due to 0 weight or missing D: 0
Finished reading data file
Default number of cross-validations: 10
Input 1 to accept the default, 2 to change it ([1:2], <cr>=1):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50):
Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3 ([1:3], <cr>=1):2
Option 2 is for equal priors.
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1):
Choose a split point selection method for numerical variables:
Choose 1 to use faster method based on sample quantiles
Choose 2 to use exhaustive search
Input 1 or 2 ([1:2], <cr>=2):
Default max. number of split levels: 10
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 2
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): hepeq.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1):
Input 1 to number all nodes, 2 to number leaves only ([1:2], <cr>=1):
Input 1 to color terminal nodes, 2 otherwise ([1:2], <cr>=1):
Choose amount of detail in nodes of LaTeX tree diagram
Input 0 for #errors, 1 for class sizes, 2 for nothing ([0:2], <cr>=1):
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values
Input your choice ([1:3], <cr>=1):3
Input file name: hepvar.txt
Contents of this file are shown below.
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: hepeq.fit
Input 2 to write R function for predicting new cases, 1 otherwise ([1:2], <cr>=1):
Input file is created!
Run GUIDE with the command: guide < hepeq.in

The resulting tree in text form is:

Node 1: ASCITES = "yes"
Node 2: die
Node 1: ASCITES /= "yes"
Node 3: SPIDERS = "no"

Wei-Yin Loh 38 GUIDE manual

4.5 Unequal misclassification costs: hepatitis data 4 CLASSIFICATION

Node 6: live
Node 3: SPIDERS /= "no"
Node 7: die

Figure 4 shows the LATEX trees using estimated priors (left) and equal priors
(right). Nodes that predict the same class have the same color. The tree using
equal priors has one more split (on SPIDERS). But both trees misclassify the same
number of samples. Therefore the left tree, being shorter, is preferred if priors are
estimated. On the other hand, since the ratio of die to live classes is 32:123,
equal priors makes each die observation equivalent to r = 123/32 = 3.84375 live
observations. Consequently, a terminal node is classified as die if its ratio of live
to die observations is less than r. Note that although only 21% of the data are
in the die class, most of these individuals are in nodes 2 and 6 (70% and 31%,
respectively).

Contents of hepvar.txt: This file summarizes the information by node:

1 c ASCITES ASCITES 1 "yes"
2 t BILIRUBIN BILIRUBIN "die"
1 c ASCITES ASCITES 1 "yes"
3 c SPIDERS SPIDERS 1 "no"
6 t MALAISE MALAISE "live"
3 c SPIDERS SPIDERS 1 "no"
7 t SEX SEX "die"

4.5 Unequal misclassification costs: hepatitis data

So far, we have assumed that the cost of misclassifying a die observation as live
is the same as the opposite. Another way to obtain a nonnull tree for the hepatitis
data is to use unequal misclassification costs. For example, if we think that the cost
of misclassifying a die observation as live is four times that of the opposite, we
will use the misclassification cost matrix
 
0 1
C=
4 0

where C(i, j) denotes the cost of classifying an observation as class i given that it
belongs to class j. Note that GUIDE sorts the class values in alphabetical order, so
that die is treated as class 1 and live as class 2 here. This matrix is saved in the
text file cost.txt which has these two lines:

Wei-Yin Loh 39 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

ASCITES ASCITES
= yes 1 =yes 1

14 18 14 SPIDERS
6 2 3 117 6 2 =no 3
die live die
5 13
88 6 7 29
live die
Figure 4: GUIDE v.26.0 0.50-SE classification tree for predicting CLASS using es-
timated (left) and equal (right) priors and unit misclassification costs. At each
split, an observation goes to the left branch if and only if the condition is satisfied.
Predicted classes (based on estimated misclassification cost) printed below terminal
nodes; sample sizes for CLASS = die and live, respectively, beside nodes. Second
best split variable at root node is SPIDERS.

0 1
4 0
The following lines in the input file generation step shows where this file is used:
Choose 1 for estimated priors, 2 for equal priors, 3 to input the priors from a file
Input 1, 2, or 3 ([1:3], <cr>=1):
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1): 2
Input the name of a file containing the cost matrix C(i|j),
where C(i|j) is the cost of classifying class j as class i
The rows of the matrix must be in alphabetical order of the class names
Input name of file: cost.txt

The resulting tree is the same as that for equal priors in Figure 4.

4.6 More than 2 classes: dermatology with ordinal predic-

tors
The data, taken from UCI (Ilter and Guvenir, 1998), give the diagnosis (6 classes)
and clinical and laboratory measurements of 34 ordinal predictor variables for 358
patients. The description and data files are derm.dsc and derm.dat, respectively.

4.6.1 Default option

The default option gives the following results.

Wei-Yin Loh 40 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Classification tree
Pruning by cross-validation
Data description file: derm.dsc
Training sample file: derm.dat
Missing value code: ?
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is class
Number of records in data file: 358
Length of longest data entry: 2
Number of classes: 6
Class proportions of dependent variable class:
Class #Cases Proportion
1 111 0.31005587
2 60 0.16759777
3 71 0.19832402
4 48 0.13407821
5 48 0.13407821
6 20 0.05586592

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 erythema s 0.0000E+00 3.0000E+00
2 scaling s 0.0000E+00 3.0000E+00
3 borders s 0.0000E+00 3.0000E+00
4 itching s 0.0000E+00 3.0000E+00
5 koebner s 0.0000E+00 3.0000E+00
6 polypap s 0.0000E+00 3.0000E+00
7 follipap s 0.0000E+00 3.0000E+00
8 oralmuc s 0.0000E+00 3.0000E+00
9 knee s 0.0000E+00 3.0000E+00
10 scalp s 0.0000E+00 3.0000E+00
11 history s 0.0000E+00 1.0000E+00
12 melanin s 0.0000E+00 3.0000E+00
13 eosin s 0.0000E+00 2.0000E+00
14 PNL s 0.0000E+00 3.0000E+00
15 fibrosis s 0.0000E+00 3.0000E+00
16 exocyto s 0.0000E+00 3.0000E+00
17 acantho s 0.0000E+00 3.0000E+00
18 hyperker s 0.0000E+00 3.0000E+00
19 paraker s 0.0000E+00 3.0000E+00
20 clubbing s 0.0000E+00 3.0000E+00
21 elongation s 0.0000E+00 3.0000E+00

Wei-Yin Loh 41 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

22 thinning s 0.0000E+00 3.0000E+00

23 spongiform s 0.0000E+00 3.0000E+00
24 munro s 0.0000E+00 3.0000E+00
25 hypergran s 0.0000E+00 3.0000E+00
26 disappea s 0.0000E+00 3.0000E+00
27 basal s 0.0000E+00 3.0000E+00
28 spongiosis s 0.0000E+00 3.0000E+00
29 sawtooth s 0.0000E+00 3.0000E+00
30 hornplug s 0.0000E+00 3.0000E+00
31 perifoll s 0.0000E+00 3.0000E+00
32 inflamm s 0.0000E+00 3.0000E+00
33 bandlike s 0.0000E+00 3.0000E+00
34 age s 0.0000E+00 7.5000E+01
35 class d 6

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
358 0 0 0 0 0 34 0 0
No. cases used for training: 358

Univariate split highest priority

Interaction and linear splits 2nd and 3rd priorities
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Simple node models
Estimated priors
Unit misclassification costs
Split values for N and S variables based on exhaustive search
Max. number of split levels: 10
Min. node sample size: 2
Number of SEs for pruned tree: 5.0000E-01

Size and CV mean cost and SE of subtrees:

Tree #Tnodes Mean Cost SE(Mean) BSE(Mean) Median Cost BSE(Median)
1 12 3.631E-02 9.887E-03 9.649E-03 2.778E-02 7.256E-03
2 10 3.631E-02 9.887E-03 9.649E-03 2.778E-02 7.256E-03
3** 9 3.631E-02 9.887E-03 9.649E-03 2.778E-02 7.256E-03
4++ 8 5.307E-02 1.185E-02 1.713E-02 2.778E-02 1.611E-02
5 7 5.866E-02 1.242E-02 1.720E-02 4.167E-02 2.080E-02
6 5 1.788E-01 2.025E-02 2.914E-02 1.690E-01 2.707E-02
7 2 4.804E-01 2.641E-02 2.058E-02 4.722E-01 2.831E-02
8 1 6.899E-01 2.444E-02 2.184E-02 6.806E-01 2.412E-02

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **

Wei-Yin Loh 42 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Selected-SE tree based on mean using bootstrap SE is marked with --

Selected-SE tree based on median and bootstrap SE is marked with ++
** tree same as -- tree
+ tree same as ++ tree
* tree same as ** tree
* tree same as -- tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Node cost is node misclassification cost divided by number of training cases

Node Total Train Predicted Node Split Interacting
label cases cases class cost variables variable
1 358 358 1 6.899E-01 polypap
2 290 290 1 6.172E-01 oralmuc
4 287 287 1 6.132E-01 fibrosis
8 239 239 1 5.356E-01 spongiosis
16 121 121 1 8.264E-02 elongation
32T 9 9 6 4.444E-01 follipap
33T 112 112 1 8.929E-03 -
17 118 118 2 5.169E-01 perifoll
34 103 103 2 4.563E-01 koebner
68 64 64 2 1.406E-01 disappea
136T 58 58 2 5.172E-02 spongiosis
137T 6 6 4 0.000E+00 -
69T 39 39 4 2.564E-02 -
35T 15 15 6 6.667E-02 -
9T 48 48 5 0.000E+00 -
5T 3 3 3 0.000E+00 -
3T 68 68 3 0.000E+00 -

Number of terminal nodes of final tree: 9

Total number of nodes of final tree: 17
Second best split variable (based on curvature test) at root node is bandlike

Classification tree:

Node 1: polypap <= 0.50000 or ?

Node 2: oralmuc <= 0.50000 or ?
Node 4: fibrosis <= 0.50000 or ?
Node 8: spongiosis <= 0.50000 or ?
Node 16: elongation <= 0.50000
Node 32: 6
Node 16: elongation > 0.50000 or ?
Node 33: 1

Wei-Yin Loh 43 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Node 8: spongiosis > 0.50000

Node 17: perifoll <= 0.50000 or ?
Node 34: koebner <= 0.50000 or ?
Node 68: disappea <= 0.50000 or ?
Node 136: 2
Node 68: disappea > 0.50000
Node 137: 4
Node 34: koebner > 0.50000
Node 69: 4
Node 17: perifoll > 0.50000
Node 35: 6
Node 4: fibrosis > 0.50000
Node 9: 5
Node 2: oralmuc > 0.50000
Node 5: 3
Node 1: polypap > 0.50000
Node 3: 3

***************************************************************

Node 1: Intermediate node

A case goes into Node 2 if polypap <= 5.0000000E-01 or ?
polypap mean = 4.4972E-01
Class Number ClassPrior
1 111 0.31006
2 60 0.16760
3 71 0.19832
4 48 0.13408
5 48 0.13408
6 20 0.05587
Number of training cases misclassified = 247
Predicted class is 1
----------------------------
Node 2: Intermediate node
A case goes into Node 4 if oralmuc <= 5.0000000E-01 or ?
oralmuc mean = 2.0690E-02
Class Number ClassPrior
1 111 0.38276
2 60 0.20690
3 3 0.01034
4 48 0.16552
5 48 0.16552
6 20 0.06897
Number of training cases misclassified = 179
Predicted class is 1

Wei-Yin Loh 44 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

----------------------------
Node 4: Intermediate node
A case goes into Node 8 if fibrosis <= 5.0000000E-01 or ?
fibrosis mean = 3.7979E-01
Class Number ClassPrior
1 111 0.38676
2 60 0.20906
3 0 0.00000
4 48 0.16725
5 48 0.16725
6 20 0.06969
Number of training cases misclassified = 176
Predicted class is 1
----------------------------
Node 8: Intermediate node
A case goes into Node 16 if spongiosis <= 5.0000000E-01 or ?
spongiosis mean = 1.0544E+00
Class Number ClassPrior
1 111 0.46444
2 60 0.25105
3 0 0.00000
4 48 0.20084
5 0 0.00000
6 20 0.08368
Number of training cases misclassified = 128
Predicted class is 1
----------------------------
Node 16: Intermediate node
A case goes into Node 32 if elongation <= 5.0000000E-01
elongation mean = 2.0909E+00
Class Number ClassPrior
1 111 0.91736
2 3 0.02479
3 0 0.00000
4 1 0.00826
5 0 0.00000
6 6 0.04959
Number of training cases misclassified = 10
Predicted class is 1
----------------------------
Node 32: Terminal node
Class Number ClassPrior
1 0 0.00000
2 3 0.33333
3 0 0.00000
4 1 0.11111

Wei-Yin Loh 45 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

5 0 0.00000
6 5 0.55556
Number of training cases misclassified = 4
Predicted class is 6
----------------------------
Node 33: Terminal node
Class Number ClassPrior
1 111 0.99107
2 0 0.00000
3 0 0.00000
4 0 0.00000
5 0 0.00000
6 1 0.00893
Number of training cases misclassified = 1
Predicted class is 1
----------------------------
Node 17: Intermediate node
A case goes into Node 34 if perifoll <= 5.0000000E-01 or ?
perifoll mean = 2.6271E-01
Class Number ClassPrior
1 0 0.00000
2 57 0.48305
3 0 0.00000
4 47 0.39831
5 0 0.00000
6 14 0.11864
Number of training cases misclassified = 61
Predicted class is 2
----------------------------
Node 34: Intermediate node
A case goes into Node 68 if koebner <= 5.0000000E-01 or ?
koebner mean = 5.4369E-01
Class Number ClassPrior
1 0 0.00000
2 56 0.54369
3 0 0.00000
4 47 0.45631
5 0 0.00000
6 0 0.00000
Number of training cases misclassified = 47
Predicted class is 2
----------------------------
Node 68: Intermediate node
A case goes into Node 136 if disappea <= 5.0000000E-01 or ?
disappea mean = 9.3750E-02
Class Number ClassPrior

Wei-Yin Loh 46 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

1 0 0.00000
2 55 0.85938
3 0 0.00000
4 9 0.14063
5 0 0.00000
6 0 0.00000
Number of training cases misclassified = 9
Predicted class is 2
----------------------------
Node 136: Terminal node
Class Number ClassPrior
1 0 0.00000
2 55 0.94828
3 0 0.00000
4 3 0.05172
5 0 0.00000
6 0 0.00000
Number of training cases misclassified = 3
Predicted class is 2
----------------------------
Node 137: Terminal node
Class Number ClassPrior
1 0 0.00000
2 0 0.00000
3 0 0.00000
4 6 1.00000
5 0 0.00000
6 0 0.00000
Number of training cases misclassified = 0
Predicted class is 4
----------------------------
Node 69: Terminal node
Class Number ClassPrior
1 0 0.00000
2 1 0.02564
3 0 0.00000
4 38 0.97436
5 0 0.00000
6 0 0.00000
Number of training cases misclassified = 1
Predicted class is 4
----------------------------
Node 35: Terminal node
Class Number ClassPrior
1 0 0.00000
2 1 0.06667

Wei-Yin Loh 47 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

3 0 0.00000
4 0 0.00000
5 0 0.00000
6 14 0.93333
Number of training cases misclassified = 1
Predicted class is 6
----------------------------
Node 9: Terminal node
Class Number ClassPrior
1 0 0.00000
2 0 0.00000
3 0 0.00000
4 0 0.00000
5 48 1.00000
6 0 0.00000
Number of training cases misclassified = 0
Predicted class is 5
----------------------------
Node 5: Terminal node
Class Number ClassPrior
1 0 0.00000
2 0 0.00000
3 3 1.00000
4 0 0.00000
5 0 0.00000
6 0 0.00000
Number of training cases misclassified = 0
Predicted class is 3
----------------------------
Node 3: Terminal node
Class Number ClassPrior
1 0 0.00000
2 0 0.00000
3 68 1.00000
4 0 0.00000
5 0 0.00000
6 0 0.00000
Number of training cases misclassified = 0
Predicted class is 3
----------------------------

Classification matrix for training sample:

Predicted True class
class 1 2 3 4 5 6
1 111 0 0 0 0 1

Wei-Yin Loh 48 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

2 0 55 0 3 0 0
3 0 0 71 0 0 0
4 0 1 0 44 0 0
5 0 0 0 0 48 0
6 0 4 0 1 0 19
Total 111 60 71 48 48 20

Number of cases used for tree construction: 358

Number misclassified: 10
Resubstitution est. of mean misclassification cost: 0.2793E-01

Observed and fitted values are stored in derm.fit

LaTeX code for tree is in derm.tex

The tree is shown in Figure 5; it misclassifies 10 observations.

4.6.2 Nearest-neighbor option

One way to obtain a smaller tree is to fit a classification model to the data in each
node and use it to classify the individual observations there. GUIDE has two means
to achieve this: nearest-neighbor and kernel discrimination. For nearest-neighbor, an
observation in a node is classified to the plurality class among observations within its
neighborhood. The neighborhood is defined to be the whole node if the split variable
is categorical. The input file for this option is obtained as follows.

Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: nn.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: nn.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1): 2
Input 1 for simple, 2 for nearest-neighbor, 3 for kernel method ([1:3], <cr>=1): 2
Choose nearest-neighbor option here.

Wei-Yin Loh 49 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

polypap
.5000 1

oralmuc
.5000 2 3 0/68
3
fibrosis
.5000 4 5 0/3
3
spongiosis
.5000 8 9 0/48
5
elongation perifoll
.5000 16 .5000 17

koebner
4/9 32 33 1/112 .5000 34 35 1/15
6 1 6
disappea
.5000 68 69 1/39
4

3/58 136 137 0/6

2 4
Figure 5: GUIDE v.26.0 0.50-SE classification tree for predicting class using
estimated priors and unit misclassification costs. At each split, an observation goes
to the left branch if and only if the condition is satisfied. The symbol stands for
or missing. Predicted classes (based on estimated misclassification cost) printed
below terminal nodes; #misclassified/sample size beside each node. Second best split
variable at root node is bandlike.

Wei-Yin Loh 50 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Input 1 for univariate, 2 for bivariate preference ([1:2], <cr>=1):

Default is univariate kernels.
Input 1 for interaction tests, 2 to skip them ([1:2], <cr>=1):
Input 1 to prune by CV, 2 by test sample, 3 for no pruning ([1:3], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: derm.dsc
Reading data description file ...
Training sample file: derm.dat
Missing value code: ?
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is class
Reading data file ...
Number of records in data file: 358
Length of longest data entry: 2
Checking for missing values ...
Total number of cases: 358
Number of classes: 6
Re-checking data ...
Assigning codes to categorical and missing values
Data checks complete
Rereading data
Class #Cases Proportion
1 111 0.31005587
2 60 0.16759777
3 71 0.19832402
4 48 0.13407821
5 48 0.13407821
6 20 0.05586592
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
358 0 0 0 0 0 34 0 0
No. cases used for training: 358
Finished reading data file
Default number of cross-validations: 10
Input 1 to accept the default, 2 to change it ([1:2], <cr>=1):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50):
Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3 ([1:3], <cr>=1):
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1):
Choose a split point selection method for numerical variables:
Choose 1 to use faster method based on sample quantiles

Wei-Yin Loh 51 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Choose 2 to use exhaustive search

Input 1 or 2 ([1:2], <cr>=2):
Default max. number of split levels: 10
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 10
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): nn.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1):
Input 1 to number all nodes, 2 to number leaves only ([1:2], <cr>=1):
Input 1 to color terminal nodes, 2 otherwise ([1:2], <cr>=1):
Choose amount of detail in nodes of LaTeX tree diagram
Input 0 for #errors, 1 for class sizes, 2 for nothing ([0:2], <cr>=0):
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values
Input your choice ([1:3], <cr>=1):
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: nn.fit
Input file is created!
Run GUIDE with the command: guide < nn.in

Results
Classification tree
Pruning by cross-validation
Data description file: derm.dsc
Training sample file: derm.dat
Missing value code: ?
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is class
Number of records in data file: 358
Length of longest data entry: 2
Number of classes: 6
Class proportions of dependent variable class:
Class #Cases Proportion
1 111 0.31005587
2 60 0.16759777
3 71 0.19832402
4 48 0.13407821
5 48 0.13407821
6 20 0.05586592

Summary information (without x variables)

Wei-Yin Loh 52 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

d=dependent, b=split and fit cat variable using 0-1 dummies,

c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 erythema s 0.0000E+00 3.0000E+00
2 scaling s 0.0000E+00 3.0000E+00
3 borders s 0.0000E+00 3.0000E+00
4 itching s 0.0000E+00 3.0000E+00
5 koebner s 0.0000E+00 3.0000E+00
6 polypap s 0.0000E+00 3.0000E+00
7 follipap s 0.0000E+00 3.0000E+00
8 oralmuc s 0.0000E+00 3.0000E+00
9 knee s 0.0000E+00 3.0000E+00
10 scalp s 0.0000E+00 3.0000E+00
11 history s 0.0000E+00 1.0000E+00
12 melanin s 0.0000E+00 3.0000E+00
13 eosin s 0.0000E+00 2.0000E+00
14 PNL s 0.0000E+00 3.0000E+00
15 fibrosis s 0.0000E+00 3.0000E+00
16 exocyto s 0.0000E+00 3.0000E+00
17 acantho s 0.0000E+00 3.0000E+00
18 hyperker s 0.0000E+00 3.0000E+00
19 paraker s 0.0000E+00 3.0000E+00
20 clubbing s 0.0000E+00 3.0000E+00
21 elongation s 0.0000E+00 3.0000E+00
22 thinning s 0.0000E+00 3.0000E+00
23 spongiform s 0.0000E+00 3.0000E+00
24 munro s 0.0000E+00 3.0000E+00
25 hypergran s 0.0000E+00 3.0000E+00
26 disappea s 0.0000E+00 3.0000E+00
27 basal s 0.0000E+00 3.0000E+00
28 spongiosis s 0.0000E+00 3.0000E+00
29 sawtooth s 0.0000E+00 3.0000E+00
30 hornplug s 0.0000E+00 3.0000E+00
31 perifoll s 0.0000E+00 3.0000E+00
32 inflamm s 0.0000E+00 3.0000E+00
33 bandlike s 0.0000E+00 3.0000E+00
34 age s 0.0000E+00 7.5000E+01
35 class d 6

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
358 0 0 0 0 0 34 0 0
No. cases used for training: 358

Univariate split highest priority

Wei-Yin Loh 53 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Interaction splits 2nd priority; no linear splits

Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Nearest-neighbor node models
Univariate preference
Estimated priors
Unit misclassification costs
Split values for N and S variables based on exhaustive search
Max. number of split levels: 10
Min. node sample size: 10
Number of SEs for pruned tree: 5.0000E-01

Size and CV mean cost and SE of subtrees:

Tree #Tnodes Mean Cost SE(Mean) BSE(Mean) Median Cost BSE(Median)
1 10 6.704E-02 1.322E-02 1.870E-02 5.556E-02 1.466E-02
2 9 6.704E-02 1.322E-02 1.870E-02 5.556E-02 1.466E-02
3 8 6.704E-02 1.322E-02 1.870E-02 5.556E-02 1.466E-02
4 7 6.704E-02 1.322E-02 1.870E-02 5.556E-02 1.466E-02
5 6 6.704E-02 1.322E-02 1.870E-02 5.556E-02 1.466E-02
6** 5 6.704E-02 1.322E-02 1.870E-02 5.556E-02 1.466E-02
7 3 1.927E-01 2.085E-02 2.111E-02 1.829E-01 2.672E-02
8 1 5.000E-01 2.643E-02 2.346E-02 5.000E-01 3.233E-02

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree and ++ tree are the same

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Node cost is node misclassification cost divided by number of training cases

Node Total Train Predicted Node Split variable followed by
label cases cases class cost (+)fit variable(s)
1 358 358 1 5.028E-01 polypap +polypap
2 290 290 1 4.517E-01 fibrosis +fibrosis
4 242 242 1 2.851E-01 spongiosis +spongiosis
8T 123 123 1 5.691E-02 elongation +elongation
9 119 119 2 4.202E-01 follipap +follipap
18T 104 104 2 1.058E-01 koebner +koebner
19T 15 15 6 6.667E-02 -
5T 48 48 5 0.000E+00 -
3T 68 68 3 0.000E+00 -

Wei-Yin Loh 54 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Number of terminal nodes of final tree: 5

Total number of nodes of final tree: 9
Second best split variable (based on curvature test) at root node is bandlike

Classification tree:

Node 1: polypap <= 0.50000 or ?

Node 2: fibrosis <= 0.50000 or ?
Node 4: spongiosis <= 0.50000 or ?
Node 8: Mean cost = 5.69106E-02
Node 4: spongiosis > 0.50000
Node 9: follipap <= 0.50000 or ?
Node 18: Mean cost = 1.05769E-01
Node 9: follipap > 0.50000
Node 19: Mean cost = 6.66667E-02
Node 2: fibrosis > 0.50000
Node 5: Mean cost = 0.00000E+00
Node 1: polypap > 0.50000
Node 3: Mean cost = 0.00000E+00

***************************************************************

Node 1: Intermediate node

A case goes into Node 2 if polypap <= 5.0000000E-01 or ?
Nearest-neighbor K = 6

polypap mean = 4.4972E-01

Fit variable
Class Number ClassPrior polypap
1 111 0.31006
2 60 0.16760
3 71 0.19832
4 48 0.13408
5 48 0.13408
6 20 0.05587
Number of training cases misclassified = 180
If node model is inapplicable due to missing values, predicted class =
1
----------------------------
Node 2: Intermediate node
A case goes into Node 4 if fibrosis <= 5.0000000E-01 or ?
Nearest-neighbor K = 6

fibrosis mean = 3.7586E-01

Wei-Yin Loh 55 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Fit variable
Class Number ClassPrior fibrosis
1 111 0.38276
2 60 0.20690
3 3 0.01034
4 48 0.16552
5 48 0.16552
6 20 0.06897
Number of training cases misclassified = 131
If node model is inapplicable due to missing values, predicted class =
1
----------------------------
Node 4: Intermediate node
A case goes into Node 8 if spongiosis <= 5.0000000E-01 or ?
Nearest-neighbor K = 6

spongiosis mean = 1.0537E+00

Fit variable
Class Number ClassPrior spongiosis
1 111 0.45868
2 60 0.24793
3 3 0.01240
4 48 0.19835
5 0 0.00000
6 20 0.08264
Number of training cases misclassified = 69
If node model is inapplicable due to missing values, predicted class =
1
----------------------------
Node 8: Terminal node
Nearest-neighbor K = 5
elongation mean = 2.0569E+00
Fit variable
Class Number ClassPrior elongation
1 111 0.90244
2 3 0.02439
3 2 0.01626
4 1 0.00813
5 0 0.00000
6 6 0.04878
----------------------------
Node 9: Intermediate node
A case goes into Node 18 if follipap <= 5.0000000E-01 or ?
Nearest-neighbor K = 5

follipap mean = 2.5210E-01

Wei-Yin Loh 56 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Fit variable
Class Number ClassPrior follipap
1 0 0.00000
2 57 0.47899
3 1 0.00840
4 47 0.39496
5 0 0.00000
6 14 0.11765
Number of training cases misclassified = 50
If node model is inapplicable due to missing values, predicted class =
2
----------------------------
Node 18: Terminal node
Nearest-neighbor K = 5
koebner mean = 5.3846E-01
Fit variable
Class Number ClassPrior koebner
1 0 0.00000
2 56 0.53846
3 1 0.00962
4 47 0.45192
5 0 0.00000
6 0 0.00000
----------------------------
Node 19: Terminal node
Nearest-neighbor K = 3
Class Number ClassPrior
1 0 0.00000
2 1 0.06667
3 0 0.00000
4 0 0.00000
5 0 0.00000
6 14 0.93333
----------------------------
Node 5: Terminal node
Nearest-neighbor K = 4
Class Number ClassPrior
1 0 0.00000
2 0 0.00000
3 0 0.00000
4 0 0.00000
5 48 1.00000
6 0 0.00000
----------------------------
Node 3: Terminal node
Nearest-neighbor K = 5

Wei-Yin Loh 57 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Class Number ClassPrior

1 0 0.00000
2 0 0.00000
3 68 1.00000
4 0 0.00000
5 0 0.00000
6 0 0.00000
----------------------------

Classification matrix for training sample:

Predicted True class
class 1 2 3 4 5 6
1 111 0 0 0 0 1
2 0 55 1 9 0 0
3 0 0 68 0 0 0
4 0 1 0 38 0 0
5 0 0 0 0 48 0
6 0 4 2 1 0 19
Total 111 60 71 48 48 20

Number of cases used for tree construction: 358

Number misclassified: 19
Resubstitution est. of mean misclassification cost: 0.5307E-01

Observed and fitted values are stored in nn.fit

LaTeX code for tree is in nn.tex

The tree is shown in Figure 6. Although it is shorter, it misclassifies 6 more

observations than if the default option is used. Note that the observations in each
terminal node are not necessarily predicted to belong to the same class.

4.6.3 Kernel density option

Another alternative is kernel discrimination models, where classification is based
on maximum likelihood with class densities estimated by the kernel method. Un-
like nearest-neighbor, however, this option also yields an estimated class probability
vector for each observation. Therefore it can serve as a nonparametric alternative
to multinomial logistic regression. Empirical evidence indicates that the nearest-
neighbor and kernel methods possess similar prediction accuracy. See Loh (2009)
for more details. Following is a log of the input file generation step for the kernel
method.

Wei-Yin Loh 58 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

polypap
.5000 1

fibrosis
.5000 2 3 0/68
3
spongiosis
.5000 4 5 0/48
5
follipap
7/123 8 .5000 9
1

11/104 18 19 1/15
2 6
Figure 6: GUIDE v.26.0 0.50-SE classification tree for predicting class using univari-
ate nearest-neighbor node models, estimated priors and unit misclassification costs.
At each split, an observation goes to the left branch if and only if the condition is
satisfied. The symbol stands for or missing. Predicted classes (based on es-
timated misclassification cost) printed below terminal nodes; #misclassified/sample
size beside each node. Second best split variable at root node is bandlike.

Wei-Yin Loh 59 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: ker.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: ker.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1): 2
Input 1 for simple, 2 for nearest-neighbor, 3 for kernel method ([1:3], <cr>=1): 3
This is where kernel density estimation is chosen.
Input 1 for univariate, 2 for bivariate preference ([1:2], <cr>=1):
Input 1 for interaction tests, 2 to skip them ([1:2], <cr>=1):
Input 1 to prune by CV, 2 by test sample, 3 for no pruning ([1:3], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: derm.dsc
Reading data description file ...
Training sample file: derm.dat
Missing value code: ?
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is class
Reading data file ...
Number of records in data file: 358
Length of longest data entry: 2
Checking for missing values ...
Total number of cases: 358
Number of classes = 6
Re-checking data ...
Assigning codes to categorical and missing values
Finished checking data
Rereading data
Class #Cases Proportion
1 111 0.31005587
2 60 0.16759777
3 71 0.19832402
4 48 0.13407821
5 48 0.13407821
6 20 0.05586592
Total #cases w/ #missing

Wei-Yin Loh 60 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
358 0 0 0 0 0 34 0 0
No. cases used for training: 358
Finished reading data file
Default number of cross-validations = 10
Input 1 to accept the default, 2 to change it ([1:2], <cr>=1):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50):
Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3 ([1:3], <cr>=1):
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1):
Choose a split point selection method for numerical variables:
Choose 1 to use faster method based on sample quantiles
Choose 2 to use exhaustive search
Input 1 or 2 ([1:2], <cr>=2):
Default max number of split levels = 10
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 10
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): ker.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1):
Input 1 to number all nodes, 2 to number leaves only ([1:2], <cr>=1):
Input 1 to color terminal nodes, 2 otherwise ([1:2], <cr>=1):
Choose amount of detail in nodes of LaTeX tree diagram
Input 0 for #errors, 1 for class sizes, 2 for nothing ([0:2], <cr>=0):
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values
Input your choice ([1:3], <cr>=1):
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: ker.fit
Input 2 to save terminal node IDs for importance scoring; 1 otherwise ([1:2], <cr>=1):
Input name of file to store predicted class and probability: ker.pro
This file contains the estimated class probabilities for each observation.
Input file is created!
Run GUIDE with the command: guide < ker.in

Results
Classification tree
Pruning by cross-validation
Data description file: derm.dsc

Wei-Yin Loh 61 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Training sample file: derm.dat

Missing value code: ?
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is class
Number of records in data file: 358
Length of longest data entry: 2
Number of classes: 6
Class proportions of dependent variable class:
Class #Cases Proportion
1 111 0.31005587
2 60 0.16759777
3 71 0.19832402
4 48 0.13407821
5 48 0.13407821
6 20 0.05586592

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 erythema s 0.0000E+00 3.0000E+00
2 scaling s 0.0000E+00 3.0000E+00
3 borders s 0.0000E+00 3.0000E+00
4 itching s 0.0000E+00 3.0000E+00
5 koebner s 0.0000E+00 3.0000E+00
6 polypap s 0.0000E+00 3.0000E+00
7 follipap s 0.0000E+00 3.0000E+00
8 oralmuc s 0.0000E+00 3.0000E+00
9 knee s 0.0000E+00 3.0000E+00
10 scalp s 0.0000E+00 3.0000E+00
11 history s 0.0000E+00 1.0000E+00
12 melanin s 0.0000E+00 3.0000E+00
13 eosin s 0.0000E+00 2.0000E+00
14 PNL s 0.0000E+00 3.0000E+00
15 fibrosis s 0.0000E+00 3.0000E+00
16 exocyto s 0.0000E+00 3.0000E+00
17 acantho s 0.0000E+00 3.0000E+00
18 hyperker s 0.0000E+00 3.0000E+00
19 paraker s 0.0000E+00 3.0000E+00
20 clubbing s 0.0000E+00 3.0000E+00
21 elongation s 0.0000E+00 3.0000E+00
22 thinning s 0.0000E+00 3.0000E+00
23 spongiform s 0.0000E+00 3.0000E+00
24 munro s 0.0000E+00 3.0000E+00

Wei-Yin Loh 62 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

25 hypergran s 0.0000E+00 3.0000E+00

26 disappea s 0.0000E+00 3.0000E+00
27 basal s 0.0000E+00 3.0000E+00
28 spongiosis s 0.0000E+00 3.0000E+00
29 sawtooth s 0.0000E+00 3.0000E+00
30 hornplug s 0.0000E+00 3.0000E+00
31 perifoll s 0.0000E+00 3.0000E+00
32 inflamm s 0.0000E+00 3.0000E+00
33 bandlike s 0.0000E+00 3.0000E+00
34 age s 0.0000E+00 7.5000E+01
35 class d 6

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
358 0 0 0 0 0 34 0 0
No. cases used for training: 358

Univariate split highest priority

Interaction splits 2nd priority; no linear splits
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Kernel density node models
Univariate preference
Estimated priors
Unit misclassification costs
Split values for N and S variables based on exhaustive search
Max. number of split levels: 10
Min. node sample size: 10
Number of SEs for pruned tree: 5.0000E-01

Size and CV mean cost and SE of subtrees:

Tree #Tnodes Mean Cost SE(Mean) BSE(Mean) Median Cost BSE(Median)
1 10 6.983E-02 1.347E-02 1.884E-02 5.556E-02 1.830E-02
2 9 6.983E-02 1.347E-02 1.884E-02 5.556E-02 1.830E-02
3 8 6.983E-02 1.347E-02 1.884E-02 5.556E-02 1.830E-02
4 7 6.983E-02 1.347E-02 1.884E-02 5.556E-02 1.830E-02
5* 6 6.983E-02 1.347E-02 1.884E-02 5.556E-02 1.830E-02
6** 5 7.263E-02 1.372E-02 1.824E-02 5.556E-02 1.440E-02
7 3 2.039E-01 2.129E-02 2.928E-02 1.829E-01 2.672E-02
8 1 5.056E-01 2.642E-02 2.881E-02 5.000E-01 3.310E-02

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++

Wei-Yin Loh 63 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

** tree same as ++ tree

** tree same as + tree
** tree same as -- tree
++ tree same as -- tree
+ tree same as ++ tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Node cost is node misclassification cost divided by number of training cases

Node Total Train Predicted Node Split variable followed by
label cases cases class cost (+)fit variable(s)
1 358 358 1 5.000E-01 polypap +polypap
2 290 290 1 4.517E-01 fibrosis +fibrosis
4 242 242 1 2.851E-01 spongiosis +spongiosis
8T 123 123 1 7.317E-02 elongation +elongation
9 119 119 2 4.874E-01 follipap +follipap
18T 104 104 2 1.058E-01 koebner +koebner
19T 15 15 6 6.667E-02 -
5T 48 48 5 0.000E+00 -
3T 68 68 3 0.000E+00 -
Split variable refers to the variable selected to split the node
and fit variable(s) refers to the one(s) used to estimate the class kernel
densities. Fit variables are indicated with a preceding + sign.
If a categorical variable is selected for fitting, discrete kernel
density estimates are used. A dash (-) indicates that a node is not split, usually
due to sample size being too small, in which case all the observations in the node
are predicted as belonging to the class that minimizes the misclassification cost.

Number of terminal nodes of final tree: 5

Total number of nodes of final tree: 9
Second best split variable (based on curvature test) at root node is bandlike

Classification tree:

Node 1: polypap <= 0.50000 or ?

Node 2: fibrosis <= 0.50000 or ?
Node 4: spongiosis <= 0.50000 or ?
Node 8: Mean cost = 7.31707E-02
Node 4: spongiosis > 0.50000
Node 9: follipap <= 0.50000 or ?
Node 18: Mean cost = 1.05769E-01
Node 9: follipap > 0.50000
Node 19: Mean cost = 6.66667E-02
Node 2: fibrosis > 0.50000

Wei-Yin Loh 64 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Node 5: Mean cost = 0.00000E+00

Node 1: polypap > 0.50000
Node 3: Mean cost = 0.00000E+00

***************************************************************

Node 1: Intermediate node

A case goes into Node 2 if polypap <= 5.0000000E-01 or ?
polypap mean = 4.4972E-01
Bandwidth
Class Number ClassPrior polypap
1 111 0.31006 3.6127E-02
2 60 0.16760 4.0857E-02
3 71 0.19832 3.9504E-01
4 48 0.13408 4.2722E-02
5 48 0.13408 4.2722E-02
6 20 0.05587 5.0897E-02
Number of training cases misclassified = 179
If node model is inapplicable due to missing values, predicted class =
1
The numbers in the last column give the kernel density bandwidth for
each class.
Number of training cases misclassified = 177
If node model is inapplicable due to missing values, predicted class =
1
----------------------------
Node 2: Intermediate node
A case goes into Node 4 if fibrosis <= 5.0000000E-01 or ?
fibrosis mean = 3.7586E-01
Bandwidth
Class Number ClassPrior fibrosis
1 111 0.38276 3.6127E-02
2 60 0.20690 4.0857E-02
3 3 0.01034 7.4383E-02
4 48 0.16552 4.2722E-02
5 48 0.16552 4.2722E-01
6 20 0.06897 5.0897E-02
Number of training cases misclassified = 131
If node model is inapplicable due to missing values, predicted class =
1
----------------------------
Node 4: Intermediate node
A case goes into Node 8 if spongiosis <= 5.0000000E-01 or ?
spongiosis mean = 1.0537E+00
Bandwidth

Wei-Yin Loh 65 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

Class Number ClassPrior spongiosis

1 111 0.45868 7.6519E-02
2 60 0.24793 4.0857E-01
3 3 0.01240 2.2315E+00
4 48 0.19835 7.5190E-01
5 0 0.00000 0.0000E+00
6 20 0.08264 1.3804E+00
Number of training cases misclassified = 69
If node model is inapplicable due to missing values, predicted class =
1
----------------------------
Node 8: Terminal node
elongation mean = 2.0569E+00
Bandwidth
Class Number ClassPrior elongation
1 111 0.90244 3.6127E-01
2 3 0.02439 7.8156E-02
3 2 0.01626 8.4758E-02
4 1 0.00813 9.7362E-02
5 0 0.00000 0.0000E+00
6 6 0.04878 7.1324E-01
----------------------------
Node 9: Intermediate node
A case goes into Node 18 if follipap <= 5.0000000E-01 or ?
follipap mean = 2.5210E-01
Bandwidth
Class Number ClassPrior follipap
1 0 0.00000 0.0000E+00
2 57 0.47899 1.4751E-01
3 1 0.00840 9.3523E-02
4 47 0.39496 4.3301E-02
5 0 0.00000 0.0000E+00
6 14 0.11765 9.0804E-01
Number of training cases misclassified = 58
If node model is inapplicable due to missing values, predicted class =
2
----------------------------
Node 18: Terminal node
koebner mean = 5.3846E-01
Bandwidth
Class Number ClassPrior koebner
1 0 0.00000 0.0000E+00
2 56 0.53846 2.9870E-01
3 1 0.00962 1.2607E-01
4 47 0.45192 8.5804E-01
5 0 0.00000 0.0000E+00

Wei-Yin Loh 66 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

6 0 0.00000 0.0000E+00
----------------------------
Node 19: Terminal node
Class Number ClassPrior
1 0 0.00000
2 1 0.06667
3 0 0.00000
4 0 0.00000
5 0 0.00000
6 14 0.93333
----------------------------
Node 5: Terminal node
Class Number ClassPrior
1 0 0.00000
2 0 0.00000
3 0 0.00000
4 0 0.00000
5 48 1.00000
6 0 0.00000
----------------------------
Node 3: Terminal node
Class Number ClassPrior
1 0 0.00000
2 0 0.00000
3 68 1.00000
4 0 0.00000
5 0 0.00000
6 0 0.00000
----------------------------

Classification matrix for training sample:

Predicted True class
class 1 2 3 4 5 6
1 111 0 0 0 0 1
2 0 58 3 10 0 5
3 0 0 68 0 0 0
4 0 1 0 38 0 0
5 0 0 0 0 48 0
6 0 1 0 0 0 14
Total 111 60 71 48 48 20

Number of cases used for tree construction: 358

Number misclassified: 21
Resubstitution est. of mean misclassification cost: 0.5866E-01

Wei-Yin Loh 67 GUIDE manual

4.6 More than 2 classes: dermatology 4 CLASSIFICATION

polypap
.5000 1

fibrosis
.5000 2 3 0/68
3
spongiosis
.5000 4 5 0/48
5
follipap
9/123 8 .5000 9
1

11/104 18 19 1/15
2 6
Figure 7: GUIDE v.26.0 0.50-SE classification tree for predicting class using uni-
variate kernel discrimination node models, estimated priors and unit misclassification
costs. At each split, an observation goes to the left branch if and only if the condition
is satisfied. The symbol stands for or missing. Predicted classes (based on es-
timated misclassification cost) printed below terminal nodes; #misclassified/sample
size beside each node. Second best split variable at root node is bandlike.

Predicted class probability estimates are stored in ker.pro

Observed and fitted values are stored in ker.fit
LaTeX code for tree is in ker.tex

The tree is shown in Figure 7. Unline the nearest-neighbor option, the kernel
option can provide an estimated class probability vector for each observation. These
are contained in the file ker.pro, the top few lines of which are given below. For
example, the probabilities that the 1st observation belongs to classes 16 are (0,
0.876, 0, 0.239, 0, 0). The last two columns give the predicted and observed class of
the observation.
"1" "2" "3" "4" "5" "6" predicted observed
0.00000 0.84423 0.03637 0.11940 0.00000 0.00000 "2" "2"
0.99616 0.00000 0.00000 0.00000 0.00000 0.00384 "1" "1"
0.00000 0.00000 1.00000 0.00000 0.00000 0.00000 "3" "3"
0.99616 0.00000 0.00000 0.00000 0.00000 0.00384 "1" "1"
0.00000 0.00000 1.00000 0.00000 0.00000 0.00000 "3" "3"

Wei-Yin Loh 68 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

4.7 More than 2 classes: heart disease, categorical predic-

tors and nodes without labels
CART and algorithms derived from it tend to be overly aggressive in their search
for splits. As a consequence, they have two significant weaknesses: (i) bias towards
selecting variables that allow more splits and (ii) long computational times when
there are categorical predictor variables with many categorical levels. These problems
are demonstrated by the heart disease data in the file heartdata.txt. The GUIDE
description file is heartdsc.txt and the class variable is num, an integer-valued code
(04) denoting a diagnosis of heart disease. There are 52 predictor variables, of which
29 are ordinal and 23 are categorical. Among the latter are the ekgmo and ekgday,
the month and day of the EKG, with 12 and 31 categorical levels, respectively. The
number of records is 617. They are obtained by combining the Hungarian, Long-
beach and Switzerland datasets from the UCI (Ilter and Guvenir, 1998) database of
the same name.

4.7.1 Input file creation

If a tree is quite large, as will be seen below, it is often preferable not to number the
nodes of the tree. The following dialog shows how to do this.

0. Read the warranty disclaimer

1. Create an input file for model fitting, importance scoring or data formatting
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: heartin.txt
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: heartout.txt
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1): 2
Option 2 allows node labels to be omitted.
Input 1 for simple, 2 for nearest-neighbor, 3 for kernel method ([1:3], <cr>=1):
Input 0 for linear, interaction and univariate splits (in this order),
1 for univariate, linear and interaction splits (in this order),
2 to skip linear splits,
3 to skip linear and interaction splits:

Wei-Yin Loh 69 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Input your choice ([0:3], <cr>=1):

Input 1 to prune by CV, 2 by test sample, 3 for no pruning ([1:3], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: heartdsc.txt
Reading data description file ...
Training sample file: heartdata.txt
Missing value code: NA
Records in data file start on line 2
Warning: N variables changed to S
Dependent variable is num
Reading data file ...
Number of records in data file: 617
Length of longest data entry: 9
Checking for missing values ...
Total number of cases: 617
Missing values found among categorical variables
Separate categories will be created for missing categorical variables
Number of classes: 5
Col. no. Categorical variable #levels #missing values
2 sex 2 0
3 painloc 2 0
4 painexer 2 0
5 relrest 2 4
6 cp 4 0
9 smoke 2 387
12 fbs 2 90
13 dm 2 545
14 famhist 2 422
15 restecg 3 2
16 ekgmo 12 53
17 ekgday 31 54
19 dig 2 66
20 prop 3 64
21 nitr 2 63
22 pro 2 61
23 diuretic 2 80
24 proto 14 112
33 exang 2 55
34 xhypo 2 58
36 slope 4 308
40 thal 7 475
53 database 3 0
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values

Wei-Yin Loh 70 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Data checks complete

Creating missing value indicators
Rereading data
Class #Cases Proportion
0 247 0.40032415
1 141 0.22852512
2 99 0.16045381
3 100 0.16207455
4 30 0.04862237
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
617 0 615 0 0 0 29 0 23
No. cases used for training: 617
No. cases excluded due to 0 weight or missing D: 0
Finished reading data file
Default number of cross-validations: 10
Input 1 to accept the default, 2 to change it ([1:2], <cr>=1):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50):
Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3 ([1:3], <cr>=1):
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1):
Choose a split point selection method for numerical variables:
Choose 1 to use faster method based on sample quantiles
Choose 2 to use exhaustive search
Input 1 or 2 ([1:2], <cr>=2):
Default max. number of split levels: 10
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 3
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): heart.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1): 2
This is where node labels are omitted.
Input 1 to color terminal nodes, 2 otherwise ([1:2], <cr>=1):
Choose amount of detail in nodes of LaTeX tree diagram
Input 0 for #errors, 1 for class sizes, 2 for nothing ([0:2], <cr>=0): 1
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values
Input your choice ([1:3], <cr>=1):
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: heart.fit
Input 2 to write R function for predicting new cases, 1 otherwise ([1:2], <cr>=1):

Wei-Yin Loh 71 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Input file is created!

Run GUIDE with the command: guide < heartin.txt

4.7.2 GUIDE results

The GUIDE tree is shown in Figure 8 and the text output follows. Despite the tree
being quite large, no categorical variable is selected to split the nodes.

Wei-Yin Loh 72 GUIDE manual

Wei-Yin Loh lmt

4.7 More than 2 classes: heart disease

.5000

rcaprox lmt
1.50 1.50
0 1
2 0
8 0
rcadist 22 rcaprox 0
1.50 0 1.50 30
30 4
3
7
cxmain 5 rcadist ladprox
1.50 0 1.50 1.50
2
73

ladprox laddist ladprox cxmain laddist cxmain

1.50 1.50 1.50 1.50 1.50 1.50
0 0 0 0 0 0
13 9 0 0 0 0
2 2 7 0 3 0
laddist 0 0 0 laddist cxmain ladprox 11 cxmain 15 ramus 27
1.50 0 0 0 1.50 1.50 1.50 0 1.50 0 1.50 0
188 10 1 2 0 0 30 0 30 30
2 6 0 7 0 27 0 0

4 CLASSIFICATION
0 0 10 2 2 2 18 0
0 0 cxmain cxmain om1 0 0 4 0 0 om1 9
0 0 1.50 ramus 1.50 0 0 0 0 0 1.50 0
0 1 58 0 0 0 0 20 1 3 1 2 0 30
GUIDE manual

5 20 15 0 32 0 0 0
0 0 1 7 4 5 19 0
0 0 0 0 0 0 0 7
0 0 0 0 0 0 0 0
0 1 1 2 1 2 2 3
Figure 8: GUIDE v.26.0 0.50-SE classification tree for predicting num using estimated priors and unit misclas-
sification costs. At each split, an observation goes to the left branch if and only if the condition is satisfied.
The symbol stands for or missing. Predicted classes (based on estimated misclassification cost)
printed below terminal nodes; sample sizes for num = 0, 1, 2, 3, and 4, respectively, beside nodes. Second
best split variable at root node is rcaprox.
4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Classification tree
Pruning by cross-validation
Data description file: heartdsc.txt
Training sample file: heartdata.txt
Missing value code: NA
Records in data file start on line 2
Warning: N variables changed to S
Dependent variable is num
Number of records in data file: 617
Length of longest data entry: 9
Missing values found among categorical variables
Separate categories will be created for missing categorical variables
Number of classes: 5
Class proportions of dependent variable num:
Class #Cases Proportion
0 247 0.40032415
1 141 0.22852512
2 99 0.16045381
3 100 0.16207455
4 30 0.04862237

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 age s 2.8000E+01 7.7000E+01
2 sex c 2
3 painloc c 2
4 painexer c 2
5 relrest c 2 4
6 cp c 4
7 trestbps s 0.0000E+00 2.0000E+02 59
8 chol s 0.0000E+00 6.0300E+02 30
9 smoke c 2 387
10 cigs s 0.0000E+00 8.0000E+01 415
11 years s 0.0000E+00 6.0000E+01 427
12 fbs c 2 90
13 dm c 2 545
14 famhist c 2 422
15 restecg c 3 2
16 ekgmo c 12 53
17 ekgday c 31 54
18 ekgyr s 8.1000E+01 8.7000E+01 53
19 dig c 2 66
20 prop c 3 64

Wei-Yin Loh 74 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

21 nitr c 2 63
22 pro c 2 61
23 diuretic c 2 80
24 proto c 14 112
25 thaldur s 1.0000E+00 2.4000E+01 56
26 thaltime s 0.0000E+00 2.0000E+01 384
27 met s 2.0000E+00 2.0000E+02 105
28 thalach s 6.0000E+01 1.9000E+02 55
29 thalrest s 3.7000E+01 1.3900E+02 56
30 tpeakbps s 1.0000E+02 2.4000E+02 63
31 tpeakbpd s 1.1000E+01 1.3400E+02 63
32 trestbpd s 0.0000E+00 1.2000E+02 59
33 exang c 2 55
34 xhypo c 2 58
35 oldpeak s -2.6000E+00 5.0000E+00 62
36 slope c 4 308
37 rldv5 s 2.0000E+00 3.6000E+01 143
38 rldv5e s 2.0000E+00 3.6000E+01 142
39 ca s 0.0000E+00 9.0000E+00 606
40 thal c 7 475
41 cyr s 1.0000E+00 8.7000E+01 9
42 num d 5
43 lmt s 0.0000E+00 1.6200E+02 275
44 ladprox s 1.0000E+00 2.0000E+00 236
45 laddist s 1.0000E+00 2.0000E+00 246
46 diag s 1.0000E+00 2.0000E+00 276
47 cxmain s 1.0000E+00 2.0000E+00 235
48 ramus s 1.0000E+00 2.0000E+00 285
49 om1 s 1.0000E+00 2.0000E+00 271
50 om2 s 1.0000E+00 2.0000E+00 290
51 rcaprox s 1.0000E+00 2.0000E+00 245
52 rcadist s 1.0000E+00 2.0000E+00 270
53 database c 3

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
617 0 615 0 0 0 29 0 23
No. cases used for training: 617
No. cases excluded due to 0 weight or missing D: 0

Univariate split highest priority

Interaction and linear splits 2nd and 3rd priorities
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Simple node models
Estimated priors

Wei-Yin Loh 75 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Unit misclassification costs

Split values for N and S variables based on exhaustive search
Max. number of split levels: 10
Min. node sample size: 3
Number of SEs for pruned tree: 5.0000E-01

Size and CV mean cost and SE of subtrees:

Tree #Tnodes Mean Cost SE(Mean) BSE(Mean) Median Cost BSE(Median)
1 38 1.037E-01 1.228E-02 1.463E-02 1.138E-01 1.956E-02
2 37 1.037E-01 1.228E-02 1.463E-02 1.138E-01 1.956E-02
3 36 1.037E-01 1.228E-02 1.463E-02 1.138E-01 1.956E-02
4 35 1.037E-01 1.228E-02 1.463E-02 1.138E-01 1.956E-02
5 34 1.037E-01 1.228E-02 1.463E-02 1.138E-01 1.956E-02
6 33 9.562E-02 1.184E-02 1.429E-02 9.836E-02 1.850E-02
7 32 9.562E-02 1.184E-02 1.429E-02 9.836E-02 1.850E-02
8* 30 9.562E-02 1.184E-02 1.429E-02 9.836E-02 1.850E-02
9 29 9.724E-02 1.193E-02 1.737E-02 9.016E-02 1.825E-02
10 27 9.887E-02 1.202E-02 1.693E-02 9.016E-02 1.775E-02
11 26 9.887E-02 1.202E-02 1.693E-02 9.016E-02 1.775E-02
12** 25 9.887E-02 1.202E-02 1.693E-02 9.016E-02 1.775E-02
13++ 23 1.021E-01 1.219E-02 1.256E-02 9.744E-02 1.696E-02
14 22 1.086E-01 1.253E-02 1.239E-02 1.066E-01 1.561E-02
15 21 1.313E-01 1.360E-02 1.296E-02 1.371E-01 1.758E-02
16 18 1.475E-01 1.428E-02 1.116E-02 1.452E-01 1.168E-02
17 17 1.556E-01 1.459E-02 7.121E-03 1.532E-01 1.170E-02
18 16 1.588E-01 1.472E-02 8.740E-03 1.532E-01 1.254E-02
19 13 1.750E-01 1.530E-02 6.284E-03 1.774E-01 8.259E-03
20 12 1.912E-01 1.583E-02 6.170E-03 1.869E-01 9.316E-03
21 9 2.998E-01 1.845E-02 2.622E-02 2.984E-01 3.642E-02
22 5 3.922E-01 1.966E-02 7.611E-03 3.983E-01 7.443E-03
23 4 3.922E-01 1.966E-02 7.611E-03 3.983E-01 7.443E-03
24 2 5.251E-01 2.010E-02 6.049E-03 5.246E-01 8.589E-03
25 1 5.997E-01 1.973E-02 3.358E-03 5.968E-01 5.312E-03

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree same as + tree
++ tree same as -- tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Wei-Yin Loh 76 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Node cost is node misclassification cost divided by number of training cases

Node Total Train Predicted Node Split Interacting
label cases cases class cost variables variable
1 617 617 0 5.997E-01 lmt
2 276 276 0 3.188E-01 rcaprox
4 244 244 0 2.295E-01 rcadist
8 229 229 0 1.790E-01 cxmain
16 211 211 0 1.090E-01 ladprox
32 196 196 0 4.082E-02 laddist
64T 190 190 0 1.053E-02 om1
65T 6 6 1 0.000E+00 -
33T 15 15 1 1.333E-01 trestbps +tpeakbpd
17 18 18 2 5.000E-01 laddist
34T 11 11 1 1.818E-01 trestbpd
35T 7 7 2 0.000E+00 -
9T 15 15 2 5.333E-01 -
5T 32 32 3 3.125E-01 trestbps :fbs
3 341 341 1 6.891E-01 lmt
6 310 310 1 6.581E-01 rcaprox
12 183 183 1 5.683E-01 rcadist
24 157 157 1 5.414E-01 ladprox
48 106 106 0 4.528E-01 laddist
96 83 83 0 3.012E-01 cxmain
192T 63 63 0 7.937E-02 om1
193T 20 20 1 0.000E+00 -
97 23 23 1 3.478E-01 cxmain +ramus
194T 16 16 1 6.250E-02 -
195T 7 7 2 0.000E+00 -
49 51 51 1 3.725E-01 cxmain
98 41 41 1 2.195E-01 om1
196T 36 36 1 1.111E-01 om2
197T 5 5 2 0.000E+00 -
99T 10 10 2 0.000E+00 -
25 26 26 3 4.231E-01 cxmain
50 15 15 1 5.333E-01 ladprox
100T 9 9 1 2.222E-01 tpeakbps
101T 6 6 3 3.333E-01 trestbpd
51T 11 11 3 0.000E+00 -
13 127 127 3 5.433E-01 ladprox
26 65 65 1 5.846E-01 laddist
52 47 47 1 4.255E-01 cxmain
104T 29 29 1 6.897E-02 thalach
105T 18 18 2 0.000E+00 -
53T 18 18 3 1.667E-01 cxmain
27 62 62 3 3.065E-01 cxmain
54 35 35 2 4.571E-01 ramus

Wei-Yin Loh 77 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

108 26 26 2 2.692E-01 om1

216T 19 19 2 0.000E+00 -
217T 7 7 3 0.000E+00 -
109T 9 9 3 0.000E+00 -
55T 27 27 3 0.000E+00 -
7T 31 31 4 3.226E-02 -

Number of terminal nodes of final tree: 25

Total number of nodes of final tree: 49
Second best split variable (based on curvature test) at root node is rcaprox

Classification tree:

Node 1: lmt <= 0.50000 or NA

Node 2: rcaprox <= 1.50000 or NA
Node 4: rcadist <= 1.50000 or NA
Node 8: cxmain <= 1.50000 or NA
Node 16: ladprox <= 1.50000 or NA
Node 32: laddist <= 1.50000 or NA
Node 64: 0
Node 32: laddist > 1.50000
Node 65: 1
Node 16: ladprox > 1.50000
Node 33: 1
Node 8: cxmain > 1.50000
Node 17: laddist <= 1.50000 or NA
Node 34: 1
Node 17: laddist > 1.50000
Node 35: 2
Node 4: rcadist > 1.50000
Node 9: 2
Node 2: rcaprox > 1.50000
Node 5: 3
Node 1: lmt > 0.50000
Node 3: lmt <= 1.50000
Node 6: rcaprox <= 1.50000 or NA
Node 12: rcadist <= 1.50000 or NA
Node 24: ladprox <= 1.50000 or NA
Node 48: laddist <= 1.50000 or NA
Node 96: cxmain <= 1.50000 or NA
Node 192: 0
Node 96: cxmain > 1.50000
Node 193: 1
Node 48: laddist > 1.50000
Node 97: 1.0000000E+00 * ramus + cxmain <= 2.5000000E+00 or NA
Node 194: 1

Wei-Yin Loh 78 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Node 97: 1.0000000E+00 * ramus + cxmain > 2.5000000E+00

Node 195: 2
Node 24: ladprox > 1.50000
Node 49: cxmain <= 1.50000 or NA
Node 98: om1 <= 1.50000 or NA
Node 196: 1
Node 98: om1 > 1.50000
Node 197: 2
Node 49: cxmain > 1.50000
Node 99: 2
Node 12: rcadist > 1.50000
Node 25: cxmain <= 1.50000
Node 50: ladprox <= 1.50000 or NA
Node 100: 1
Node 50: ladprox > 1.50000
Node 101: 3
Node 25: cxmain > 1.50000 or NA
Node 51: 3
Node 6: rcaprox > 1.50000
Node 13: ladprox <= 1.50000 or NA
Node 26: laddist <= 1.50000 or NA
Node 52: cxmain <= 1.50000 or NA
Node 104: 1
Node 52: cxmain > 1.50000
Node 105: 2
Node 26: laddist > 1.50000
Node 53: 3
Node 13: ladprox > 1.50000
Node 27: cxmain <= 1.50000 or NA
Node 54: ramus <= 1.50000 or NA
Node 108: om1 <= 1.50000 or NA
Node 216: 2
Node 108: om1 > 1.50000
Node 217: 3
Node 54: ramus > 1.50000
Node 109: 3
Node 27: cxmain > 1.50000
Node 55: 3
Node 3: lmt > 1.50000
Node 7: 4

***************************************************************

In the following the predictor node mean is mean of complete cases.

Node 1: Intermediate node

Wei-Yin Loh 79 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

A case goes into Node 2 if lmt <= 5.0000000E-01 or NA

lmt mean = 1.5556E+00
Class Number ClassPrior
0 247 0.40032
1 141 0.22853
2 99 0.16045
3 100 0.16207
4 30 0.04862
Number of training cases misclassified = 370
Predicted class is 0
----------------------------
Node 2: Intermediate node
A case goes into Node 4 if rcaprox <= 1.5000000E+00 or NA
rcaprox mean = 1.7619E+00
Class Number ClassPrior
0 188 0.68116
1 35 0.12681
2 26 0.09420
3 27 0.09783
4 0 0.00000
Number of training cases misclassified = 88
Predicted class is 0
----------------------------
Node 4: Intermediate node
A case goes into Node 8 if rcadist <= 1.5000000E+00 or NA
rcadist mean = 1.7895E+00
Class Number ClassPrior
0 188 0.77049
1 33 0.13525
2 18 0.07377
3 5 0.02049
4 0 0.00000
Number of training cases misclassified = 56
Predicted class is 0
----------------------------
Node 8: Intermediate node
A case goes into Node 16 if cxmain <= 1.5000000E+00 or NA
cxmain mean = 1.6923E+00
Class Number ClassPrior
0 188 0.82096
1 30 0.13100
2 11 0.04803
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 41
Predicted class is 0

Wei-Yin Loh 80 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

----------------------------
Node 16: Intermediate node
A case goes into Node 32 if ladprox <= 1.5000000E+00 or NA
ladprox mean = 1.6818E+00
Class Number ClassPrior
0 188 0.89100
1 21 0.09953
2 2 0.00948
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 23
Predicted class is 0
----------------------------
Node 32: Intermediate node
A case goes into Node 64 if laddist <= 1.5000000E+00 or NA
laddist mean = 1.2857E+00
Class Number ClassPrior
0 188 0.95918
1 8 0.04082
2 0 0.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 8
Predicted class is 0
----------------------------
Node 64: Terminal node
Class Number ClassPrior
0 188 0.98947
1 2 0.01053
2 0 0.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 2
Predicted class is 0
----------------------------
Node 65: Terminal node
Class Number ClassPrior
0 0 0.00000
1 6 1.00000
2 0 0.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 0
Predicted class is 1
----------------------------
Node 33: Terminal node

Wei-Yin Loh 81 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Class Number ClassPrior

0 0 0.00000
1 13 0.86667
2 2 0.13333
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 2
Predicted class is 1
----------------------------
Node 17: Intermediate node
A case goes into Node 34 if laddist <= 1.5000000E+00 or NA
laddist mean = 1.8750E+00
Class Number ClassPrior
0 0 0.00000
1 9 0.50000
2 9 0.50000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 9
Predicted class is 2
----------------------------
Node 34: Terminal node
Class Number ClassPrior
0 0 0.00000
1 9 0.81818
2 2 0.18182
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 2
Predicted class is 1
----------------------------
Node 35: Terminal node
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 7 1.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 0
Predicted class is 2
----------------------------
Node 9: Terminal node
Class Number ClassPrior
0 0 0.00000
1 3 0.20000
2 7 0.46667

Wei-Yin Loh 82 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

3 5 0.33333
4 0 0.00000
Number of training cases misclassified = 8
Predicted class is 2
----------------------------
Node 5: Terminal node
Class Number ClassPrior
0 0 0.00000
1 2 0.06250
2 8 0.25000
3 22 0.68750
4 0 0.00000
Number of training cases misclassified = 10
Predicted class is 3
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if lmt <= 1.5000000E+00
lmt mean = 1.5601E+00
Class Number ClassPrior
0 59 0.17302
1 106 0.31085
2 73 0.21408
3 73 0.21408
4 30 0.08798
Number of training cases misclassified = 235
Predicted class is 1
----------------------------
Node 6: Intermediate node
A case goes into Node 12 if rcaprox <= 1.5000000E+00 or NA
rcaprox mean = 1.4137E+00
Class Number ClassPrior
0 58 0.18710
1 106 0.34194
2 73 0.23548
3 73 0.23548
4 0 0.00000
Number of training cases misclassified = 204
Predicted class is 1
----------------------------
Node 12: Intermediate node
A case goes into Node 24 if rcadist <= 1.5000000E+00 or NA
rcadist mean = 1.1444E+00
Class Number ClassPrior
0 58 0.31694
1 79 0.43169
2 31 0.16940

Wei-Yin Loh 83 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

3 15 0.08197
4 0 0.00000
Number of training cases misclassified = 104
Predicted class is 1
----------------------------
Node 24: Intermediate node
A case goes into Node 48 if ladprox <= 1.5000000E+00 or NA
ladprox mean = 1.3290E+00
Class Number ClassPrior
0 58 0.36943
1 72 0.45860
2 27 0.17197
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 85
Predicted class is 1
----------------------------
Node 48: Intermediate node
A case goes into Node 96 if laddist <= 1.5000000E+00 or NA
laddist mean = 1.2212E+00
Class Number ClassPrior
0 58 0.54717
1 40 0.37736
2 8 0.07547
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 48
Predicted class is 0
----------------------------
Node 96: Intermediate node
A case goes into Node 192 if cxmain <= 1.5000000E+00 or NA
cxmain mean = 1.2439E+00
Class Number ClassPrior
0 58 0.69880
1 25 0.30120
2 0 0.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 25
Predicted class is 0
----------------------------
Node 192: Terminal node
Class Number ClassPrior
0 58 0.92063
1 5 0.07937
2 0 0.00000

Wei-Yin Loh 84 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 5
Predicted class is 0
----------------------------
Node 193: Terminal node
Class Number ClassPrior
0 0 0.00000
1 20 1.00000
2 0 0.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 0
Predicted class is 1
----------------------------
Node 97: Intermediate node
A case goes into Node 194 if
1.0000000E+00 * ramus + cxmain <= 2.5000000E+00 or NA
Linear combination mean = 2.3043E+00
Class Number ClassPrior
0 0 0.00000
1 15 0.65217
2 8 0.34783
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 8
Predicted class is 1
----------------------------
Node 194: Terminal node
Class Number ClassPrior
0 0 0.00000
1 15 0.93750
2 1 0.06250
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 1
Predicted class is 1
----------------------------
Node 195: Terminal node
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 7 1.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 0

Wei-Yin Loh 85 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Predicted class is 2
----------------------------
Node 49: Intermediate node
A case goes into Node 98 if cxmain <= 1.5000000E+00 or NA
cxmain mean = 1.2000E+00
Class Number ClassPrior
0 0 0.00000
1 32 0.62745
2 19 0.37255
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 19
Predicted class is 1
----------------------------
Node 98: Intermediate node
A case goes into Node 196 if om1 <= 1.5000000E+00 or NA
om1 mean = 1.1250E+00
Class Number ClassPrior
0 0 0.00000
1 32 0.78049
2 9 0.21951
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 9
Predicted class is 1
----------------------------
Node 196: Terminal node
Class Number ClassPrior
0 0 0.00000
1 32 0.88889
2 4 0.11111
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 4
Predicted class is 1
----------------------------
Node 197: Terminal node
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 5 1.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 0
Predicted class is 2
----------------------------

Wei-Yin Loh 86 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Node 99: Terminal node

Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 10 1.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 0
Predicted class is 2
----------------------------
Node 25: Intermediate node
A case goes into Node 50 if cxmain <= 1.5000000E+00
cxmain mean = 1.4000E+00
Class Number ClassPrior
0 0 0.00000
1 7 0.26923
2 4 0.15385
3 15 0.57692
4 0 0.00000
Number of training cases misclassified = 11
Predicted class is 3
----------------------------
Node 50: Intermediate node
A case goes into Node 100 if ladprox <= 1.5000000E+00 or NA
ladprox mean = 1.4000E+00
Class Number ClassPrior
0 0 0.00000
1 7 0.46667
2 4 0.26667
3 4 0.26667
4 0 0.00000
Number of training cases misclassified = 8
Predicted class is 1
----------------------------
Node 100: Terminal node
Class Number ClassPrior
0 0 0.00000
1 7 0.77778
2 2 0.22222
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 2
Predicted class is 1
----------------------------
Node 101: Terminal node
Class Number ClassPrior

Wei-Yin Loh 87 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

0 0 0.00000
1 0 0.00000
2 2 0.33333
3 4 0.66667
4 0 0.00000
Number of training cases misclassified = 2
Predicted class is 3
----------------------------
Node 51: Terminal node
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 0 0.00000
3 11 1.00000
4 0 0.00000
Number of training cases misclassified = 0
Predicted class is 3
----------------------------
Node 13: Intermediate node
A case goes into Node 26 if ladprox <= 1.5000000E+00 or NA
ladprox mean = 1.4882E+00
Class Number ClassPrior
0 0 0.00000
1 27 0.21260
2 42 0.33071
3 58 0.45669
4 0 0.00000
Number of training cases misclassified = 69
Predicted class is 3
----------------------------
Node 26: Intermediate node
A case goes into Node 52 if laddist <= 1.5000000E+00 or NA
laddist mean = 1.2769E+00
Class Number ClassPrior
0 0 0.00000
1 27 0.41538
2 23 0.35385
3 15 0.23077
4 0 0.00000
Number of training cases misclassified = 38
Predicted class is 1
----------------------------
Node 52: Intermediate node
A case goes into Node 104 if cxmain <= 1.5000000E+00 or NA
cxmain mean = 1.3830E+00
Class Number ClassPrior

Wei-Yin Loh 88 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

0 0 0.00000
1 27 0.57447
2 20 0.42553
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 20
Predicted class is 1
----------------------------
Node 104: Terminal node
Class Number ClassPrior
0 0 0.00000
1 27 0.93103
2 2 0.06897
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 2
Predicted class is 1
----------------------------
Node 105: Terminal node
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 18 1.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 0
Predicted class is 2
----------------------------
Node 53: Terminal node
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 3 0.16667
3 15 0.83333
4 0 0.00000
Number of training cases misclassified = 3
Predicted class is 3
----------------------------
Node 27: Intermediate node
A case goes into Node 54 if cxmain <= 1.5000000E+00 or NA
cxmain mean = 1.4355E+00
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 19 0.30645
3 43 0.69355

Wei-Yin Loh 89 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

4 0 0.00000
Number of training cases misclassified = 19
Predicted class is 3
----------------------------
Node 54: Intermediate node
A case goes into Node 108 if ramus <= 1.5000000E+00 or NA
ramus mean = 1.2571E+00
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 19 0.54286
3 16 0.45714
4 0 0.00000
Number of training cases misclassified = 16
Predicted class is 2
----------------------------
Node 108: Intermediate node
A case goes into Node 216 if om1 <= 1.5000000E+00 or NA
om1 mean = 1.2692E+00
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 19 0.73077
3 7 0.26923
4 0 0.00000
Number of training cases misclassified = 7
Predicted class is 2
----------------------------
Node 216: Terminal node
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 19 1.00000
3 0 0.00000
4 0 0.00000
Number of training cases misclassified = 0
Predicted class is 2
----------------------------
Node 217: Terminal node
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 0 0.00000
3 7 1.00000
4 0 0.00000
Number of training cases misclassified = 0

Wei-Yin Loh 90 GUIDE manual

4.7 More than 2 classes: heart disease 4 CLASSIFICATION

Predicted class is 3
----------------------------
Node 109: Terminal node
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 0 0.00000
3 9 1.00000
4 0 0.00000
Number of training cases misclassified = 0
Predicted class is 3
----------------------------
Node 55: Terminal node
Class Number ClassPrior
0 0 0.00000
1 0 0.00000
2 0 0.00000
3 27 1.00000
4 0 0.00000
Number of training cases misclassified = 0
Predicted class is 3
----------------------------
Node 7: Terminal node
Class Number ClassPrior
0 1 0.03226
1 0 0.00000
2 0 0.00000
3 0 0.00000
4 30 0.96774
Number of training cases misclassified = 1
Predicted class is 4
----------------------------

Classification matrix for training sample:

Predicted True class
class 0 1 2 3 4
0 246 7 0 0 0
1 0 129 13 0 0
2 0 3 73 5 0
3 0 2 13 95 0
4 1 0 0 0 30
Total 247 141 99 100 30

Number of cases used for tree construction: 617

Number misclassified: 44

Wei-Yin Loh 91 GUIDE manual

 5 REGRESSION

painexer=a
0|
247/141/99/100/30

database=a lmt< 1.5

0 1
rcaprox< 1.5
0 177/30/17/22/5 0 70/111/82/78/25
143/9/3/7/2 34/21/14/15/3
ladprox< 1.5
0 3 rcaprox< 1.5 database=a
34/17/6/4/1 0/4/8/11/2 1 0
cxmain< 1.5
0 1 25/85/59/58/0 45/26/23/20/25
0 1 0/8/5/2/0
34/9/1/2/1 oldpeak< 0.9
0 4
34/4/1/0/1
0/5/0/2/0 44/26/23/20/13 1/0/0/0/12
ekgday=djlmprvwCD
0 1
ekgday=afghijkmnorstuvwxyBDE ladprox< 1.5 34/9/7/4/2 10/17/16/16/11
1 3
25/62/25/11/0 0/23/34/47/0 ekgday=bcefhnosxzABE
1 3
1/14/3/1/2 ekgmo=bcdehkl
1 2 cxmain< 1.5 ekgday=abcilmotA 29/3/13/15/94
1 3
19/58/10/9/0
6/4/15/2/0 9/3/12/12/1 0/0/1/3/8
0/23/19/10/0 0/0/15/37/0 ekgmo=beh
2 3 3 2
laddist< 1.5 0/0/12/6/0
0/0/3/31/0 0 3 2/1/11/2/0
7/2/1/10/1
1 2 6/2/0/2/0
1/0/1/8/1
1 3 0/0/14/6/0
0/23/5/4/0
0/23/2/0/0
0/0/3/4/0

Figure 9: RPART model for heart disease data

Resubstitution est. of mean misclassification cost: 0.7131E-01

Observed and fitted values are stored in heart.fit

LaTeX code for tree is in heart.tex

4.7.3 RPART model

The GUIDE model in Figure 8 took 3 sec. to construct on a Linux computer. In
comparison, RPART (Therneau et al., 2017) took more than 3.5 hrs, due primarily
to the presence of ekgmo and ekgday. The result is shown in Figure 9. It splits
repeatedly on ekgmo and ekgday.

5 Regression
GUIDE can fit least-squares (LS), quantile, Poisson, proportional hazards, and least-
median-of-squares (LMS) regression tree models. We use the birthweight data in files
birthwt.dat and birthwt.dsc to demonstrate LS models. The data consist of ob-
servations from 50,000 live births. They are a subset of a larger dataset analyzed

Wei-Yin Loh 92 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

in Koenker and Hallock (2001); see also Koenker (2005). The variables are weight
(infant birth weight), black (indicator of black mother), married (indicator of mar-
ried mother), boy (indicator of boy), visit (prenatal visit: 0 = no visits, 1 = visit
in 2nd trimester, 2 = visit in last trimester, 3 = visit in 1st trimester), ed (Mothers
education level: 0 = high school, 1 = some college, 2 = college, 3 = less than high
school), smoke (indicator of smoking mother), cigsper (number of cigarettes smoked
per day), age (mothers age), and wtgain (mothers weight gain during pregnancy).
The contents of birthwt.dsc are:
birthwt.dat
NA
1
1 weight d
2 black c
3 married c
4 boy c
5 age n
6 smoke c
7 cigsper n
8 wtgain n
9 visit c
10 ed c
11 lowbwt x

The last variable lowbwt is a derived indicator of low birthweight not used here.

5.1 Least squares constant: birthwt data

5.1.1 Input file creation
The input file cons.in is obtained as follows. We select the non-default option to
enable more selections to be provided.
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: cons.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: cons.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2

Wei-Yin Loh 93 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

Choose type of regression model:

1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1):
Input 1 for least squares, 2 least median of squares ([1:2], <cr>=1):
Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 0 for stepwise regression in each node (R var. may not be included),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
0: stepwise linear, 1: multiple linear, 2: best polynomial, 3: constant,
4: stepwise simple ANCOVA ([0:4], <cr>=0):3
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):2
We choose 2 to allow more options below.
Input number of cross-validations ([2:50000], <cr>=10):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50):
Choose a split point selection method for numerical variables:
Choose 1 to use faster method based on sample quantiles
Choose 2 to use exhaustive search
Input 1 or 2 ([1:2], <cr>=2):
Default max. number of split levels: 30
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 250
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): cons.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1):2
Choose a color for the terminal nodes:
(1) white
(2) lightgray
(3) gray
(4) darkgray
(5) black
(6) yellow
(7) red
(8) blue
(9) green
(10) magenta
(11) cyan
Input your choice ([1:11], <cr>=6):
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values

Wei-Yin Loh 94 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

Input your choice ([1:3], <cr>=1):3

Input file name: cons.var
Choose 3 to save split variable information to a separate file.
Input file name: cons.var
Input 2 to save regressor names in a file, 1 otherwise ([1:2], <cr>=1):
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr >=2):
Input name of file to store node ID and fitted value of each case: cons.fit
Input 2 to write R function for predicting new cases, 1 otherwise ([1:2], <cr> =1):
Input file is created!
Run GUIDE with the command: guide < cons.in

5.1.2 Results
The contents of cons.out follow.
Least squares regression tree
Pruning by cross-validation
Data description file: birthwt.dsc
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is weight
Piecewise constant model
Number of records in data file: 50000
Length of longest data entry: 4

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 weight d 2.4000E+02 6.3500E+03
2 black c 2
3 married c 2
4 boy c 2
5 age s 1.8000E+01 4.5000E+01
6 smoke c 2
7 cigsper s 0.0000E+00 6.0000E+01
8 wtgain s 0.0000E+00 9.8000E+01
9 visit c 4
10 ed c 4

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var

Wei-Yin Loh 95 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

50000 0 0 1 0 0 3 0 6
No weight variable in data file
No. cases used for training: 50000

Interaction tests on all variables

Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Split values for N and S variables based on exhaustive search
Max. number of split levels: 30
Min. node sample size: 250
Number of SEs for pruned tree: 5.0000E-01

Size and CV MSE and SE of subtrees:

Tree #Tnodes Mean MSE SE(Mean) BSE(Mean) Median MSE BSE(Median)
1 147 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.163E+03
2 146 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.163E+03
3 145 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.163E+03
4 144 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.165E+03
5 143 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.165E+03
6 142 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.165E+03
7 141 2.891E+05 2.802E+03 1.698E+03 2.874E+05 2.162E+03
8 140 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.166E+03
9 139 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.166E+03
10 138 2.891E+05 2.802E+03 1.698E+03 2.873E+05 2.165E+03
11 137 2.891E+05 2.802E+03 1.698E+03 2.873E+05 2.166E+03
12 136 2.891E+05 2.802E+03 1.697E+03 2.873E+05 2.166E+03
13 135 2.891E+05 2.802E+03 1.700E+03 2.873E+05 2.160E+03
14 134 2.891E+05 2.802E+03 1.701E+03 2.873E+05 2.161E+03
15 133 2.891E+05 2.802E+03 1.700E+03 2.873E+05 2.166E+03
16 132 2.891E+05 2.802E+03 1.700E+03 2.873E+05 2.165E+03
17 131 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.162E+03
18 130 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.163E+03
19 128 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.163E+03
20 127 2.891E+05 2.802E+03 1.698E+03 2.873E+05 2.163E+03
21 125 2.891E+05 2.802E+03 1.698E+03 2.873E+05 2.164E+03
22 124 2.891E+05 2.802E+03 1.698E+03 2.874E+05 2.164E+03
23 123 2.891E+05 2.802E+03 1.697E+03 2.874E+05 2.166E+03
24 121 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.181E+03
25 120 2.891E+05 2.802E+03 1.699E+03 2.873E+05 2.181E+03
26 119 2.891E+05 2.802E+03 1.698E+03 2.874E+05 2.176E+03
27 118 2.891E+05 2.802E+03 1.700E+03 2.874E+05 2.176E+03
28 117 2.891E+05 2.802E+03 1.700E+03 2.874E+05 2.172E+03
29 116 2.891E+05 2.802E+03 1.699E+03 2.874E+05 2.171E+03
30 114 2.891E+05 2.802E+03 1.699E+03 2.874E+05 2.168E+03
31 113 2.891E+05 2.802E+03 1.699E+03 2.874E+05 2.168E+03
32 110 2.891E+05 2.802E+03 1.701E+03 2.874E+05 2.168E+03

Wei-Yin Loh 96 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

33 109 2.891E+05 2.802E+03 1.701E+03 2.874E+05 2.168E+03

34 108 2.891E+05 2.802E+03 1.701E+03 2.874E+05 2.169E+03
35 107 2.891E+05 2.802E+03 1.702E+03 2.874E+05 2.169E+03
36 106 2.891E+05 2.802E+03 1.700E+03 2.874E+05 2.167E+03
37 103 2.891E+05 2.803E+03 1.694E+03 2.874E+05 2.157E+03
38 101 2.891E+05 2.803E+03 1.694E+03 2.874E+05 2.157E+03
39 100 2.891E+05 2.803E+03 1.694E+03 2.874E+05 2.149E+03
40 99 2.891E+05 2.803E+03 1.697E+03 2.873E+05 2.156E+03
41 97 2.891E+05 2.803E+03 1.702E+03 2.873E+05 2.153E+03
42 96 2.891E+05 2.803E+03 1.702E+03 2.873E+05 2.148E+03
43 95 2.891E+05 2.803E+03 1.703E+03 2.874E+05 2.173E+03
44 94 2.891E+05 2.802E+03 1.723E+03 2.874E+05 2.185E+03
45 93 2.891E+05 2.802E+03 1.722E+03 2.874E+05 2.175E+03
46 92 2.891E+05 2.802E+03 1.722E+03 2.874E+05 2.177E+03
47 90 2.891E+05 2.802E+03 1.722E+03 2.874E+05 2.177E+03
48 89 2.891E+05 2.802E+03 1.727E+03 2.874E+05 2.175E+03
49 88 2.891E+05 2.802E+03 1.727E+03 2.874E+05 2.175E+03
50 87 2.890E+05 2.802E+03 1.722E+03 2.874E+05 2.175E+03
51 86 2.890E+05 2.802E+03 1.722E+03 2.874E+05 2.175E+03
52 85 2.891E+05 2.802E+03 1.719E+03 2.875E+05 2.167E+03
53 84 2.891E+05 2.802E+03 1.719E+03 2.875E+05 2.167E+03
54 83 2.891E+05 2.802E+03 1.719E+03 2.875E+05 2.167E+03
55 82 2.890E+05 2.802E+03 1.728E+03 2.874E+05 2.161E+03
56 80 2.890E+05 2.802E+03 1.727E+03 2.874E+05 2.152E+03
57 78 2.890E+05 2.802E+03 1.727E+03 2.874E+05 2.152E+03
58 76 2.890E+05 2.802E+03 1.726E+03 2.874E+05 2.153E+03
59 74 2.890E+05 2.802E+03 1.723E+03 2.874E+05 2.152E+03
60 73 2.890E+05 2.802E+03 1.727E+03 2.874E+05 2.172E+03
61 72 2.890E+05 2.801E+03 1.746E+03 2.874E+05 2.197E+03
62 71 2.890E+05 2.801E+03 1.746E+03 2.874E+05 2.197E+03
63 70 2.890E+05 2.801E+03 1.746E+03 2.874E+05 2.197E+03
64 69 2.890E+05 2.801E+03 1.747E+03 2.874E+05 2.204E+03
65 68 2.890E+05 2.801E+03 1.745E+03 2.874E+05 2.204E+03
66 66 2.890E+05 2.801E+03 1.747E+03 2.874E+05 2.226E+03
67 65 2.889E+05 2.800E+03 1.718E+03 2.874E+05 2.183E+03
68* 64 2.889E+05 2.800E+03 1.724E+03 2.874E+05 2.134E+03
69 63 2.889E+05 2.800E+03 1.722E+03 2.874E+05 2.134E+03
70 61 2.890E+05 2.802E+03 1.739E+03 2.874E+05 2.144E+03
71 60 2.890E+05 2.802E+03 1.739E+03 2.874E+05 2.144E+03
72 58 2.890E+05 2.802E+03 1.739E+03 2.874E+05 2.144E+03
73 57 2.890E+05 2.802E+03 1.739E+03 2.874E+05 2.144E+03
74 56 2.890E+05 2.802E+03 1.740E+03 2.874E+05 2.152E+03
75 55 2.890E+05 2.802E+03 1.740E+03 2.874E+05 2.152E+03
76 54 2.890E+05 2.802E+03 1.740E+03 2.874E+05 2.152E+03
77 53 2.890E+05 2.802E+03 1.740E+03 2.874E+05 2.152E+03
78 51 2.889E+05 2.801E+03 1.749E+03 2.874E+05 2.153E+03

Wei-Yin Loh 97 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

79 50 2.889E+05 2.801E+03 1.748E+03 2.874E+05 2.150E+03

80 49 2.889E+05 2.801E+03 1.748E+03 2.874E+05 2.150E+03
81 47 2.889E+05 2.801E+03 1.748E+03 2.874E+05 2.150E+03
82 46 2.889E+05 2.801E+03 1.748E+03 2.874E+05 2.150E+03
83 44 2.889E+05 2.801E+03 1.748E+03 2.874E+05 2.150E+03
84 42 2.889E+05 2.801E+03 1.748E+03 2.874E+05 2.150E+03
85 41 2.889E+05 2.801E+03 1.754E+03 2.872E+05 2.188E+03
86 40 2.890E+05 2.802E+03 1.737E+03 2.872E+05 2.167E+03
87 38 2.890E+05 2.802E+03 1.739E+03 2.872E+05 2.183E+03
88 37 2.890E+05 2.802E+03 1.727E+03 2.872E+05 2.166E+03
89 36 2.891E+05 2.805E+03 1.716E+03 2.873E+05 2.034E+03
90 35 2.891E+05 2.806E+03 1.714E+03 2.873E+05 2.008E+03
91 34 2.891E+05 2.806E+03 1.707E+03 2.872E+05 1.965E+03
92 33 2.890E+05 2.807E+03 1.711E+03 2.872E+05 1.968E+03
93+ 32 2.891E+05 2.811E+03 1.699E+03 2.872E+05 1.971E+03
94 31 2.892E+05 2.812E+03 1.715E+03 2.876E+05 2.041E+03
95++ 30 2.892E+05 2.812E+03 1.712E+03 2.876E+05 2.028E+03
96 29 2.894E+05 2.815E+03 1.703E+03 2.883E+05 2.214E+03
97 28 2.894E+05 2.815E+03 1.715E+03 2.883E+05 2.228E+03
98 26 2.895E+05 2.816E+03 1.773E+03 2.883E+05 2.232E+03
99-- 25 2.896E+05 2.816E+03 1.777E+03 2.883E+05 2.260E+03
100 24 2.898E+05 2.820E+03 1.711E+03 2.888E+05 2.002E+03
101 23 2.898E+05 2.820E+03 1.730E+03 2.888E+05 2.021E+03
102 22 2.898E+05 2.820E+03 1.730E+03 2.888E+05 2.021E+03
103 21 2.899E+05 2.822E+03 1.753E+03 2.888E+05 2.004E+03
104** 20 2.902E+05 2.825E+03 1.782E+03 2.888E+05 2.171E+03
105 19 2.903E+05 2.826E+03 1.750E+03 2.888E+05 2.059E+03
106 18 2.903E+05 2.826E+03 1.750E+03 2.888E+05 2.059E+03
107 16 2.912E+05 2.836E+03 1.618E+03 2.904E+05 2.047E+03
108 15 2.915E+05 2.837E+03 1.609E+03 2.908E+05 2.048E+03
109 14 2.916E+05 2.838E+03 1.592E+03 2.911E+05 2.035E+03
110 13 2.918E+05 2.838E+03 1.650E+03 2.911E+05 1.887E+03
111 12 2.919E+05 2.840E+03 1.637E+03 2.911E+05 1.767E+03
112 11 2.919E+05 2.840E+03 1.637E+03 2.911E+05 1.767E+03
113 10 2.939E+05 2.846E+03 1.677E+03 2.924E+05 1.941E+03
114 9 2.953E+05 2.855E+03 1.625E+03 2.934E+05 1.900E+03
115 8 2.959E+05 2.857E+03 1.640E+03 2.934E+05 2.011E+03
116 7 2.968E+05 2.861E+03 1.595E+03 2.956E+05 1.845E+03
117 6 2.972E+05 2.861E+03 1.562E+03 2.956E+05 2.020E+03
118 5 2.994E+05 2.866E+03 1.652E+03 2.976E+05 2.276E+03
119 4 3.029E+05 2.889E+03 2.029E+03 3.020E+05 3.119E+03
120 3 3.065E+05 2.909E+03 1.652E+03 3.059E+05 2.312E+03
121 2 3.106E+05 2.954E+03 1.517E+03 3.094E+05 1.738E+03
122 1 3.208E+05 3.107E+03 1.608E+03 3.204E+05 2.111E+03

0-SE tree based on mean is marked with *

Wei-Yin Loh 98 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

0-SE tree based on median is marked with +

Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

D-mean is mean of weight in the node

Cases fit give the number of cases used to fit node
MSE is residual sum of squares divided by number of cases in node
Node Total Cases Matrix Node Node Split Interacting
label cases fit rank D-mean MSE variable variable
1 50000 50000 1 3.371E+03 3.208E+05 wtgain
2 20241 20241 1 3.247E+03 3.707E+05 black
4 3831 3831 1 3.041E+03 4.230E+05 cigsper
8T 3425 3425 1 3.070E+03 4.155E+05 boy
9T 406 406 1 2.804E+03 4.242E+05 -
5 16410 16410 1 3.295E+03 3.463E+05 cigsper
10 13965 13965 1 3.335E+03 3.306E+05 age
20T 3234 3234 1 3.226E+03 3.288E+05 boy
21 10731 10731 1 3.368E+03 3.265E+05 boy
42T 5363 5363 1 3.320E+03 3.031E+05 age
43T 5368 5368 1 3.417E+03 3.452E+05 married
11T 2445 2445 1 3.064E+03 3.739E+05 visit
3 29759 29759 1 3.455E+03 2.693E+05 married
6 8291 8291 1 3.332E+03 2.715E+05 wtgain
12 5399 5399 1 3.280E+03 2.658E+05 boy
24 2616 2616 1 3.220E+03 2.497E+05 black
48T 909 909 1 3.131E+03 2.628E+05 ed
49 1707 1707 1 3.268E+03 2.363E+05 cigsper
98T 1262 1262 1 3.330E+03 2.202E+05 visit
99T 445 445 1 3.093E+03 2.410E+05 -
25T 2783 2783 1 3.336E+03 2.746E+05 black
13T 2892 2892 1 3.429E+03 2.676E+05 wtgain
7 21468 21468 1 3.503E+03 2.604E+05 boy
14 10148 10148 1 3.437E+03 2.425E+05 cigsper
28 9290 9290 1 3.457E+03 2.379E+05 wtgain
56T 4812 4812 1 3.406E+03 2.300E+05 black
57 4478 4478 1 3.512E+03 2.406E+05 black
114T 359 359 1 3.320E+03 2.980E+05 -
115T 4119 4119 1 3.528E+03 2.322E+05 ed
29T 858 858 1 3.224E+03 2.427E+05 wtgain
15 11320 11320 1 3.561E+03 2.692E+05 wtgain
30 6607 6607 1 3.511E+03 2.631E+05 cigsper

Wei-Yin Loh 99 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

60T 6094 6094 1 3.529E+03 2.582E+05 age

61T 513 513 1 3.288E+03 2.692E+05 age
31 4713 4713 1 3.632E+03 2.692E+05 cigsper
62 4254 4254 1 3.652E+03 2.680E+05 black
124T 336 336 1 3.451E+03 3.218E+05 -
125T 3918 3918 1 3.669E+03 2.597E+05 age
63T 459 459 1 3.450E+03 2.442E+05 -

Number of terminal nodes of final tree: 20

Total number of nodes of final tree: 39
Second best split variable (based on curvature test) at root node is married

Regression tree:
At each categorical variable split, values not in training data go right

Node 1: wtgain <= 27.50000

Node 2: black = "1"
Node 4: cigsper <= 1.50000 or NA
Node 8: weight-mean = 3.06959E+03
Node 4: cigsper > 1.50000
Node 9: weight-mean = 2.80386E+03
Node 2: black /= "1"
Node 5: cigsper <= 0.50000 or NA
Node 10: age <= 23.50000
Node 20: weight-mean = 3.22560E+03
Node 10: age > 23.50000 or NA
Node 21: boy = "0"
Node 42: weight-mean = 3.31952E+03
Node 21: boy /= "0"
Node 43: weight-mean = 3.41662E+03
Node 5: cigsper > 0.50000
Node 11: weight-mean = 3.06438E+03
Node 1: wtgain > 27.50000 or NA
Node 3: married = "0"
Node 6: wtgain <= 40.50000 or NA
Node 12: boy = "0"
Node 24: black = "1"
Node 48: weight-mean = 3.13074E+03
Node 24: black /= "1"
Node 49: cigsper <= 2.50000 or NA
Node 98: weight-mean = 3.32954E+03
Node 49: cigsper > 2.50000
Node 99: weight-mean = 3.09315E+03
Node 12: boy /= "0"
Node 25: weight-mean = 3.33626E+03
Node 6: wtgain > 40.50000

Wei-Yin Loh 100 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

Node 13: weight-mean = 3.42949E+03

Node 3: married /= "0"
Node 7: boy = "0"
Node 14: cigsper <= 0.50000 or NA
Node 28: wtgain <= 35.50000 or NA
Node 56: weight-mean = 3.40565E+03
Node 28: wtgain > 35.50000
Node 57: black = "1"
Node 114: weight-mean = 3.31965E+03
Node 57: black /= "1"
Node 115: weight-mean = 3.52837E+03
Node 14: cigsper > 0.50000
Node 29: weight-mean = 3.22371E+03
Node 7: boy /= "0"
Node 15: wtgain <= 38.50000 or NA
Node 30: cigsper <= 1.50000 or NA
Node 60: weight-mean = 3.52948E+03
Node 30: cigsper > 1.50000
Node 61: weight-mean = 3.28845E+03
Node 15: wtgain > 38.50000
Node 31: cigsper <= 0.50000 or NA
Node 62: black = "1"
Node 124: weight-mean = 3.45097E+03
Node 62: black /= "1"
Node 125: weight-mean = 3.66879E+03
Node 31: cigsper > 0.50000
Node 63: weight-mean = 3.44990E+03

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap solution
, see
Loh et al. (2016), "Identification of subgroups with differential treatment eff
ects
for longitudinal and multiresponse variables", Statistics in Medicine, v.35, 48
37-4855.

Node 1: Intermediate node

A case goes into Node 2 if wtgain <= 2.7500000E+01
wtgain mean = 3.0709E+01
Coefficients of least squares regression function:
Regressor Coefficient t-stat p-val
Constant 3.3708E+03 1330.76 0.0000
Mean of weight = 3370.75664000000
----------------------------

Wei-Yin Loh 101 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

Node 2: Intermediate node

A case goes into Node 4 if black = "1"
black mode = "0"
----------------------------
Node 4: Intermediate node
A case goes into Node 8 if cigsper <= 1.5000000E+00 or NA
cigsper mean = 9.4179E-01
----------------------------
Node 8: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.0696E+03 278.70 0.0000
Mean of weight = 3069.58919708029
----------------------------
Node 9: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 2.8039E+03 86.74 0.0000
Mean of weight = 2803.85960591133
----------------------------
Node 5: Intermediate node
A case goes into Node 10 if cigsper <= 5.0000000E-01 or NA
cigsper mean = 1.8370E+00
----------------------------
Node 10: Intermediate node
A case goes into Node 20 if age <= 2.3500000E+01
age mean = 2.8223E+01
----------------------------
Node 20: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.2256E+03 319.91 0.0000
Mean of weight = 3225.59523809524
----------------------------
Node 21: Intermediate node
A case goes into Node 42 if boy = "0"
boy mode = "1"
----------------------------
Node 42: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.3195E+03 441.53 0.0000
Mean of weight = 3319.51724780906
----------------------------
Node 43: Terminal node
Coefficients of least squares regression functions:

Wei-Yin Loh 102 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

Regressor Coefficient t-stat p-val

Constant 3.4166E+03 426.08 0.0000
Mean of weight = 3416.61997019374
----------------------------
Node 11: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.0644E+03 247.79 0.0000
Mean of weight = 3064.38445807771
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if married = "0"
married mode = "1"
----------------------------
Node 6: Intermediate node
A case goes into Node 12 if wtgain <= 4.0500000E+01 or NA
wtgain mean = 3.9897E+01
----------------------------
Node 12: Intermediate node
A case goes into Node 24 if boy = "0"
boy mode = "1"
----------------------------
Node 24: Intermediate node
A case goes into Node 48 if black = "1"
black mode = "0"
----------------------------
Node 48: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.1307E+03 184.11 0.0000
Mean of weight = 3130.74257425743
----------------------------
Node 49: Intermediate node
A case goes into Node 98 if cigsper <= 2.5000000E+00 or NA
cigsper mean = 3.2191E+00
----------------------------
Node 98: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.3295E+03 252.05 0.0000
Mean of weight = 3329.54437400951
----------------------------
Node 99: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.0931E+03 132.92 0.0000

Wei-Yin Loh 103 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

Mean of weight = 3093.14831460674

----------------------------
Node 25: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.3363E+03 335.85 0.0000
Mean of weight = 3336.26266618757
----------------------------
Node 13: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.4295E+03 356.49 0.0000
Mean of weight = 3429.48928077455
----------------------------
Node 7: Intermediate node
A case goes into Node 14 if boy = "0"
boy mode = "1"
----------------------------
Node 14: Intermediate node
A case goes into Node 28 if cigsper <= 5.0000000E-01 or NA
cigsper mean = 9.2836E-01
----------------------------
Node 28: Intermediate node
A case goes into Node 56 if wtgain <= 3.5500000E+01 or NA
wtgain mean = 3.7944E+01
----------------------------
Node 56: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.4057E+03 492.62 0.0000
Mean of weight = 3405.65378221114
----------------------------
Node 57: Intermediate node
A case goes into Node 114 if black = "1"
black mode = "0"
----------------------------
Node 114: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.3197E+03 115.22 0.0000
Mean of weight = 3319.65459610028
----------------------------
Node 115: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.5284E+03 469.96 0.0000

Wei-Yin Loh 104 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

Mean of weight = 3528.36999271668

----------------------------
Node 29: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.2237E+03 191.68 0.0000
Mean of weight = 3223.70629370629
----------------------------
Node 15: Intermediate node
A case goes into Node 30 if wtgain <= 3.8500000E+01 or NA
wtgain mean = 3.8360E+01
----------------------------
Node 30: Intermediate node
A case goes into Node 60 if cigsper <= 1.5000000E+00 or NA
cigsper mean = 9.0056E-01
----------------------------
Node 60: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.5295E+03 542.26 0.0000
Mean of weight = 3529.48194945848
----------------------------
Node 61: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.2885E+03 143.56 0.0000
Mean of weight = 3288.45419103314
----------------------------
Node 31: Intermediate node
A case goes into Node 62 if cigsper <= 5.0000000E-01 or NA
cigsper mean = 1.0269E+00
----------------------------
Node 62: Intermediate node
A case goes into Node 124 if black = "1"
black mode = "0"
----------------------------
Node 124: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.4510E+03 111.52 0.0000
Mean of weight = 3450.97023809524
----------------------------
Node 125: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.6688E+03 450.62 0.0000

Wei-Yin Loh 105 GUIDE manual

5.1 Least squares constant: birthwt data 5 REGRESSION

Mean of weight = 3668.79198570699

----------------------------
Node 63: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val
Constant 3.4499E+03 149.58 0.0000
Mean of weight = 3449.90413943355
----------------------------

Proportion of variance (R-squared) explained by tree model: .0938

Observed and fitted values are stored in cons.fit

LaTeX code for tree is in cons.tex
Split and fit variable names are stored in cons.var

Figure 10 shows the tree diagram. It is rather large and may not be so easy
to interpret. This is because the complexity of a piecewise-constant model rests
completely in the tree structure. GUIDE has 4 options that will reduce the tree size
by moving some of the complexity to the nodes of the tree.

Wei-Yin Loh 106 GUIDE manual

Wei-Yin Loh

5.1 Least squares constant: birthwt data

wtgain
27.50

black married
=1 =0

cigsper cigsper wtgain boy

1.50 .5000 40.50 =0

age boy cigsper wtgain

23.50 =0 .5000 38.50
3425 406 2445 2892
3070. 2804. 3064. 3429.
boy black wtgain cigsper cigsper
107

=0 =1 35.50 1.50 .5000

3234 2783 858
3226. 3336. 3224.
cigsper black black
2.50 =1 =1
5363 5368 909 4812 6094 513 459
3320. 3417. 3131. 3406. 3529. 3288. 3450.

1262 445 359 4119 336 3918

5 REGRESSION
3330. 3093. 3320. 3528. 3451. 3669.
GUIDE manual

Figure 10: GUIDE v.26.0 0.50-SE piecewise constant least-squares regression tree for predicting weight. At
each split, an observation goes to the left branch if and only if the condition is satisfied. The symbol
stands for or missing. For splits on categorical variables, values not present in the training sample go to
the right. Sample size (in italics) and mean of weight printed below nodes. Second best split variable at
root node is married.
5.1 Least squares constant: birthwt data 5 REGRESSION

5.1.3 Contents of cons.var

Following are the contents of the text file cons.var. Column 1 gives the node
number, column 2 is a c, s, or t, depending on whether the split variable is C or S,
or if the node is terminal. Column 3 gives the name of the split variable; if the node
is terminal, the name is printed as NONE. Column 4 gives the name of the interacting
variable if it is present; if there is no interacting variable, the split variable name
is repeated. If a node is nonterminal, column 5 contains an integer indicating the
number of parameter values to follow on the same line. For example, the integer is 1
for node 1 and it is followed by the value 0.2750000000E+02 which is the split point.
(If a split is on a categorical variable, column 5 will give the number of categorical
values defined by the split and subsequent columns will give those values.) If a node
is terminal, column 5 gives the node mean of the D variable. The main purpose of
this file is to facilitate machine extraction of the split information without parsing
cons.out.
1 s wtgain wtgain 1 0.2750000000E+02
2 c black black 1 "1"
4 s cigsper cigsper 1 0.1500000000E+01
8 t boy boy 0.3069589197E+04
9 t NONE NONE 0.2803859606E+04
2 c black black 1 "1"
5 s cigsper cigsper 1 0.5000000000E+00
10 s age age 1 0.2350000000E+02
20 t boy boy 0.3225595238E+04
21 c boy boy 1 "0"
42 t age age 0.3319517248E+04
21 c boy boy 1 "0"
43 t married married 0.3416619970E+04
11 t visit visit 0.3064384458E+04
3 c married married 1 "0"
6 s wtgain wtgain 1 0.4050000000E+02
12 c boy boy 1 "0"
24 c black black 1 "1"
48 t ed ed 0.3130742574E+04
24 c black black 1 "1"
49 s cigsper cigsper 1 0.2500000000E+01
98 t visit visit 0.3329544374E+04
99 t NONE NONE 0.3093148315E+04
12 c boy boy 1 "0"
25 t black black 0.3336262666E+04
13 t wtgain wtgain 0.3429489281E+04
3 c married married 1 "0"
7 c boy boy 1 "0"
14 s cigsper cigsper 1 0.5000000000E+00
28 s wtgain wtgain 1 0.3550000000E+02
56 t black black 0.3405653782E+04

Wei-Yin Loh 108 GUIDE manual

5.2 Least squares simple linear: birthwt data 5 REGRESSION

57 c black black 1 "1"

114 t NONE NONE 0.3319654596E+04
57 c black black 1 "1"
115 t ed ed 0.3528369993E+04
29 t wtgain wtgain 0.3223706294E+04
7 c boy boy 1 "0"
15 s wtgain wtgain 1 0.3850000000E+02
30 s cigsper cigsper 1 0.1500000000E+01
60 t age age 0.3529481949E+04
61 t age age 0.3288454191E+04
31 s cigsper cigsper 1 0.5000000000E+00
62 c black black 1 "1"
124 t NONE NONE 0.3450970238E+04
62 c black black 1 "1"
125 t age age 0.3668791986E+04
63 t NONE NONE 0.3449904139E+04

5.2 Least squares simple linear: birthwt data

Instead of fitting a constant in each node, we can fit a simple polynomial regression
model of the form y = 0 +1 x+2 x2 +. . . , k xk , where the degree of the polynomial
k is pre-specified and the best predictor variable x is chosen at each node by GUIDE.
We demonstrate this with k = 1 here.

5.2.1 Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: lin.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: lin.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1):
Input 1 for least squares, 2 least median of squares ([1:2], <cr>=1):

Wei-Yin Loh 109 GUIDE manual

5.2 Least squares simple linear: birthwt data 5 REGRESSION

Choose complexity of model to use at each node:

If R variable present (i.e., subgroup identification),
choose 0 for stepwise regression in each node (R var. may not be included),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
0: stepwise linear, 1: multiple linear, 2: best polynomial, 3: constant,
4: stepwise simple ANCOVA ([0:4], <cr>=0):2
This option chooses the best predictor variable at each node.
Input 1 for default options, 2 otherwise ([1:2], <cr>=1): 2
Choosing 2 opens more options.
Input degree of polynomial ([1:9], <cr>=1):
Choose 1 to use alpha-level to drop insignificant powers, 2 otherwise ([1:2], <cr>=1):
Input significance level ([0.00:1.00], <cr>=0.05):
Choose a truncation method for predicted values:
0: none, 1: node range, 2: +10% node range, 3: global range,
4: 2-sided Winsorization Winsorization
Input 0, 1, 2, 3, or 4 ([0:4], <cr>=3):
Input 1 for interaction tests, 2 to skip them ([1:2], <cr>=1):
Input 1 to prune by CV, 2 for no pruning ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: birthwt.dsc
Reading data description file ...
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Reading data file ...
Number of records in data file: 50000
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 50000
Col. no. Categorical variable #levels #missing values
2 black 2 0
3 married 2 0
4 boy 2 0
6 smoke 2 0
9 visit 4 0
10 ed 4 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished processing 5000 of 50000 observations
Finished processing 10000 of 50000 observations
Finished processing 15000 of 50000 observations
Finished processing 20000 of 50000 observations

Wei-Yin Loh 110 GUIDE manual

5.2 Least squares simple linear: birthwt data 5 REGRESSION

Finished processing 25000 of 50000 observations

Finished processing 30000 of 50000 observations
Finished processing 35000 of 50000 observations
Finished processing 40000 of 50000 observations
Finished processing 45000 of 50000 observations
Finished processing 50000 of 50000 observations
Data checks complete
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 0 6
No weight variable in data file
No. cases used for training: 50000
Finished reading data file
Choose how you wish to deal with missing values in training or test data:
Option 1: Fit separate models to complete and incomplete cases
Option 2: Impute missing F and N values at each node with means for regression
Option 3: Fit a piecewise constant model
Input selection: ([1:3], <cr>=2):
Default number of cross-validations: 10
Input 1 to accept the default, 2 to change it ([1:2], <cr>=1):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50):
Choose fraction of cases for splitting
Larger values give more splits: 0 = median split and 1 = all possible splits
Default fraction is 1.0000
Choose 1 to accept default split fraction, 2 to change it
Input 1 or 2 ([1:2], <cr>=1):
Default max. number of split levels: 30
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 2499
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): lin.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1):
Input 1 to number all nodes, 2 to number leaves only ([1:2], <cr>=1):
Choose color(s) for the terminal nodes:
(1) red-yellow-green
(2) red-green-blue
(3) magenta-yellow-green
(4) yellow
(5) green
(6) magenta
(7) cyan
(8) lightgray

Wei-Yin Loh 111 GUIDE manual

5.2 Least squares simple linear: birthwt data 5 REGRESSION

(9) white
Input your choice ([1:9], <cr>=1):
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values
Input your choice ([1:3], <cr>=1):3
Choose 3 to save split variable info to a file.
Input file name: lin.var
Input 2 to save regressor names in a file, 1 otherwise ([1:2], <cr>=1): 2
Option 2 saves names of regressors and their coefs to a file.
Input file name: lin.reg
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: lin.it
Input 2 to save terminal node IDs for importance scoring; 1 otherwise ([1:2], <cr>=1):
Input 2 to write R function for predicting new cases, 1 otherwise ([1:2], <cr>=1):
Input file is created!
Run GUIDE with the command: guide < lin.in

5.2.2 Results
Warning: The p-values produced by GUIDE are not adjusted for split
selection. Therefore they are typically biased low. One way to adjust
the p-values to control for split selection is with the bootstrap method in
Loh et al. (2016). Our experience indicates, however, that any unadjusted
p-value less than 0.01 is likely to be significant at level 0.05 after the
bootstrap adjustment.
Least squares regression tree
Predictions truncated at global min and max of D sample values
The default option sets this truncation.
Pruning by cross-validation
Data description file: birthwt.dsc
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Piecewise simple linear or constant model
Powers are dropped if they are not significant at level .0500
The default option sets non-significant regression coefs to 0.
Number of records in data file: 50000
Length of longest data entry: 4

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,

Wei-Yin Loh 112 GUIDE manual

5.2 Least squares simple linear: birthwt data 5 REGRESSION

c=split-only categorical, n=split and fit numerical, f=fit-only numerical,

s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 weight d 2.4000E+02 6.3500E+03
2 black c 2
3 married c 2
4 boy c 2
5 age n 1.8000E+01 4.5000E+01
6 smoke c 2
7 cigsper n 0.0000E+00 6.0000E+01
8 wtgain n 0.0000E+00 9.8000E+01
9 visit c 4
10 ed c 4
C variables are not used as predictors in the node models.
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 0 6
No weight variable in data file
No. cases used for training: 50000

Missing values imputed with node means for model fitting

Interaction tests on all variables
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node = 1.0000
Max number of split levels = 30
Minimum node size = 2499

Number of SEs for pruned tree = 5.0000E-01

Size and CV MSE and SE of subtrees:

Tree #Tnodes Mean MSE SE(Mean) BSE(Mean) Median MSE BSE(Median)
1 13 2.909E+05 2.797E+03 1.725E+03 2.904E+05 2.205E+03
2* 12 2.909E+05 2.798E+03 1.725E+03 2.904E+05 2.199E+03
3+ 11 2.909E+05 2.797E+03 1.727E+03 2.903E+05 2.158E+03
4 10 2.910E+05 2.798E+03 1.771E+03 2.904E+05 2.304E+03
5 9 2.910E+05 2.798E+03 1.771E+03 2.904E+05 2.304E+03
6 8 2.910E+05 2.798E+03 1.771E+03 2.904E+05 2.304E+03
7-- 7 2.910E+05 2.798E+03 1.771E+03 2.904E+05 2.304E+03
8 6 2.918E+05 2.803E+03 1.820E+03 2.911E+05 2.268E+03
9 5 2.918E+05 2.803E+03 1.820E+03 2.911E+05 2.268E+03
10** 4 2.918E+05 2.803E+03 1.820E+03 2.911E+05 2.268E+03
11 3 2.944E+05 2.805E+03 1.897E+03 2.934E+05 2.381E+03
12 2 2.994E+05 2.833E+03 1.756E+03 2.989E+05 1.995E+03
13 1 3.069E+05 2.887E+03 1.553E+03 3.059E+05 1.759E+03

Wei-Yin Loh 113 GUIDE manual

5.2 Least squares simple linear: birthwt data 5 REGRESSION

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree same as ++ tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

D-mean is mean of weight in the node

Cases fit give the number of cases used to fit node
MSE and R^2 are based on all cases in node
Node Total Cases Matrix Node Node Node Split Other
label cases fit rank D-mean MSE R^2 variable variables
1 50000 50000 2 3.371E+03 3.069E+05 0.0432 married +wtgain
2T 14369 14369 2 3.234E+03 3.166E+05 0.0558 black +wtgain
3 35631 35631 2 3.426E+03 2.925E+05 0.0394 cigsper +wtgain
6 32318 32318 2 3.449E+03 2.856E+05 0.0366 boy +wtgain
12T 15610 15610 2 3.388E+03 2.681E+05 0.0336 age +wtgain
13T 16708 16708 2 3.506E+03 2.958E+05 0.0380 age +wtgain
7T 3313 3313 2 3.198E+03 3.063E+05 0.0587 - +wtgain

Number of terminal nodes of final tree: 4

Total number of nodes of final tree: 7
Second best split variable (based on curvature test) at root node is black

Regression tree:

Node 1: married = "0"

Node 2: weight-mean = 3.23443E+03
Node 1: married /= "0"
Node 3: cigsper <= 0.50000 or NA
Node 6: boy = "0"
Node 12: weight-mean = 3.38775E+03
Node 6: boy /= "0"
Node 13: weight-mean = 3.50636E+03
Node 3: cigsper > 0.50000
Node 7: weight-mean = 3.19811E+03

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap solution
, see

Wei-Yin Loh 114 GUIDE manual

5.2 Least squares simple linear: birthwt data 5 REGRESSION

Loh et al. (2016), "Identification of subgroups with differential treatment eff

ects
for longitudinal and multiresponse variables", Statistics in Medicine, v.35, 48
37-4855.

Node 1: Intermediate node

A case goes into Node 2 if married = "0"
married mode = "1"
Coefficients of least squares regression function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 3.0900E+03 482.24 0.0000
wtgain 9.1433E+00 47.52 0.0000 0.0000E+00 3.0709E+01 9.8000E+01
Mean of weight = 3370.75664000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 2: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 2.9346E+03 259.41 0.0000
wtgain 9.7545E+00 29.13 0.0000 0.0000E+00 3.0737E+01 9.8000E+01
Mean of weight = 3234.42870067506
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if cigsper <= 5.0000000E-01 or NA
cigsper mean = 1.0771E+00
----------------------------
Node 6: Intermediate node
A case goes into Node 12 if boy = "0"
boy mode = "1"
----------------------------
Node 12: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 3.1476E+03 283.21 0.0000
wtgain 7.8926E+00 23.29 0.0000 0.0000E+00 3.0429E+01 9.8000E+01
Mean of weight = 3387.74631646381
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 13: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 3.2296E+03 279.14 0.0000
wtgain 8.8854E+00 25.68 0.0000 0.0000E+00 3.1149E+01 9.8000E+01
Mean of weight = 3506.35934881494
Predicted values truncated at 240.000000000000 & 6350.00000000000

Wei-Yin Loh 115 GUIDE manual

5.2 Least squares simple linear: birthwt data 5 REGRESSION

married
=0 1

cigsper
2 .5000 3
14369
3234. boy
+wtgain =0 6 7
3313
3198.
+wtgain
12 13
15610 16708
3388. 3506.
+wtgain +wtgain
Figure 11: GUIDE v.26.0 0.50-SE piecewise simple linear least-squares regression
tree for predicting weight. At each split, an observation goes to the left branch if
and only if the condition is satisfied. The symbol stands for or missing.
Sample size (in italics), mean of weight, and sign and name of regressor variables
printed below nodes. Nodes with negative and positive slopes are colored red and
green, respectively. Second best split variable at root node is black.

----------------------------
Node 7: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 2.9060E+03 129.19 0.0000
wtgain 9.8392E+00 14.36 0.0000 0.0000E+00 2.9688E+01 9.8000E+01
Mean of weight = 3198.11469966797
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------

Proportion of variance (R-squared) explained by tree model: .0842

Observed and fitted values are stored in lin.fit

Regressor names and coefficients are stored in lin.reg
LaTeX code for tree is in lin.tex
Split and fit variable names are stored in lin.var

The tree model is shown in Figure 11. Besides being much smaller than the
piecewise-constant model, it shows that wtgain (mothers weight gain) is the best

Wei-Yin Loh 116 GUIDE manual

5.3 Multiple linear: birthwt data 5 REGRESSION

linear predictor in every node.

5.2.3 Contents of lin.var

The contents of lin.var follow. Their interpretation is the same as for the piecewise
constant model above.
1 c married married 1 "0"
2 t black black 0.3234428701E+04
1 c married married 1 "0"
3 n cigsper cigsper 1 0.5000000000E+00
6 c boy boy 1 "0"
12 t age age 0.3387746316E+04
6 c boy boy 1 "0"
13 t age age 0.3506359349E+04
7 t NONE NONE 0.3198114700E+04

5.2.4 Contents of lin.reg

The first row of the file contains the column names. Column 1 gives the node number
and column 2 the linear predictor variable in the node. Columns 3 and 4 give the
regression coefficients (intercept followed by slope) and columns 5 and 6 the lower
and upper truncation points for the predicted D values. This file is useful for machine
extraction of the regression information in each node.
node variable 0 1 lower upper
2 wtgain 2.935E+03 9.754E+00 2.400E+02 6.350E+03
12 wtgain 3.148E+03 7.893E+00 2.400E+02 6.350E+03
13 wtgain 3.230E+03 8.885E+00 2.400E+02 6.350E+03
7 wtgain 2.906E+03 9.839E+00 2.400E+02 6.350E+03

5.3 Multiple linear: birthwt data

The tree structure complexity may be reduced further by fitting a multiple linear
regression in each node as follows.

5.3.1 Input file creation

We again use non-defaults to allow more options.
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file

Wei-Yin Loh 117 GUIDE manual

5.3 Multiple linear: birthwt data 5 REGRESSION

Input your choice: 1

Name of batch input file: mul.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: mul.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1):
Input 1 for least squares, 2 least median of squares ([1:2], <cr>=1):
Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 0 for stepwise regression in each node (R var. may not be included),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
0: stepwise linear, 1: multiple linear, 2: best polynomial, 3: constant,
4: stepwise simple ANCOVA ([0:4], <cr>=0):1
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):2
More options are opened by selecting 2.
Input 2 for no intercept term, 1 otherwise ([1:2], <cr>=1):
Choose a truncation method for predicted values:
0: none, 1: node range, 2: +10% node range, 3: global range
Input 0, 1, 2, or 3 ([0:3], <cr>=3):
Input 1 for interaction tests, 2 to skip them ([1:2], <cr>=1):
Input 1 to prune by CV, 2 for no pruning ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: birthwt.dsc
Reading data description file ...
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Reading data file ...
Number of records in data file: 50000
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 50000
Col. no. Categorical variable #levels #missing values
2 black 2 0
3 married 2 0

Wei-Yin Loh 118 GUIDE manual

5.3 Multiple linear: birthwt data 5 REGRESSION

4 boy 2 0
6 smoke 2 0
9 visit 4 0
10 ed 4 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished processing 5000 of 50000 observations
Finished processing 10000 of 50000 observations
Finished processing 15000 of 50000 observations
Finished processing 20000 of 50000 observations
Finished processing 25000 of 50000 observations
Finished processing 30000 of 50000 observations
Finished processing 35000 of 50000 observations
Finished processing 40000 of 50000 observations
Finished processing 45000 of 50000 observations
Finished processing 50000 of 50000 observations
Data checks complete
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 0 6
No weight variable in data file
No. cases used for training: 50000
Finished reading data file
Choose how you wish to deal with missing values in training or test data:
Option 1: Fit separate models to complete and incomplete cases
Option 2: Impute missing F and N values at each node with means for regression
Option 3: Fit a piecewise constant model
Input selection: ([1:3], <cr>=2):
These options are irrelevant here; they deal with missing values.
Default number of cross-validations: 10
Input 1 to accept the default, 2 to change it ([1:2], <cr>=1):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50):
Choose a split point selection method for numerical variables:
Choose 1 to use faster method based on sample quantiles
Choose 2 to use exhaustive search
Input 1 or 2 ([1:2], <cr>=2):
Default max. number of split levels: 30
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 2499
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): mul.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1):

Wei-Yin Loh 119 GUIDE manual

5.3 Multiple linear: birthwt data 5 REGRESSION

Input 1 to number all nodes, 2 to number leaves only ([1:2], <cr>=1):

Choose a color for the terminal nodes:
(1) white
(2) lightgray
(3) gray
(4) darkgray
(5) black
(6) yellow
(7) red
(8) blue
(9) green
(10) magenta
(11) cyan
Input your choice ([1:11], <cr>=6):
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values
Input your choice ([1:3], <cr>=1):3
Input file name: mul.var
Input 2 to save truncation limits and regression coefficients in a file, 1 otherwise ([1:2], <cr>=1
Input file name: mul.reg
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: mul.fit
Input 2 to write R function for predicting new cases, 1 otherwise ([1:2], <cr>=1):
Input file is created!
Run GUIDE with the command: guide < mul.in

5.3.2 Results
Least squares regression tree
Predictions truncated at global min and max of D sample values
Truncation of predicted values can be changed by selecting the non-default option.
Pruning by cross-validation
Data description file: birthwt.dsc
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Piecewise linear model
Number of records in data file: 50000
Length of longest data entry: 4

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,

Wei-Yin Loh 120 GUIDE manual

5.3 Multiple linear: birthwt data 5 REGRESSION

s=split-only numerical, w=weight

Column Name Minimum Maximum #Categories #Missing
1 weight d 2.4000E+02 6.3500E+03
2 black c 2
3 married c 2
4 boy c 2
5 age n 1.8000E+01 4.5000E+01
6 smoke c 2
7 cigsper n 0.0000E+00 6.0000E+01
8 wtgain n 0.0000E+00 9.8000E+01
9 visit c 4
10 ed c 4

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 0 6
No weight variable in data file
No. cases used for training: 50000

Missing values imputed with node means for regression

Interaction tests on all variables
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Split values for N and S variables based on exhaustive search
Max. number of split levels: 30
Min. node sample size: 2499
100 bootstrap calibration replicates
Scaling for N variables after bootstrap calibration: 1.000
Number of SEs for pruned tree: 5.0000E-01

Size and CV MSE and SE of subtrees:

Tree #Tnodes Mean MSE SE(Mean) BSE(Mean) Median MSE BSE(Median)
1 13 2.880E+05 2.771E+03 1.820E+03 2.879E+05 2.004E+03
2 12 2.880E+05 2.771E+03 1.820E+03 2.879E+05 2.010E+03
3 11 2.880E+05 2.771E+03 1.823E+03 2.879E+05 2.009E+03
4 10 2.880E+05 2.772E+03 1.823E+03 2.879E+05 2.021E+03
5 9 2.880E+05 2.772E+03 1.833E+03 2.879E+05 2.015E+03
6+ 8 2.879E+05 2.772E+03 1.841E+03 2.876E+05 2.005E+03
7* 7 2.879E+05 2.770E+03 1.836E+03 2.877E+05 1.977E+03
8 6 2.879E+05 2.771E+03 1.853E+03 2.876E+05 1.992E+03
9 5 2.879E+05 2.771E+03 1.853E+03 2.876E+05 1.992E+03
10-- 4 2.885E+05 2.777E+03 1.857E+03 2.876E+05 2.001E+03
11** 3 2.888E+05 2.781E+03 1.820E+03 2.883E+05 2.001E+03
12 2 2.919E+05 2.781E+03 1.888E+03 2.905E+05 2.256E+03
13 1 2.989E+05 2.859E+03 1.653E+03 2.972E+05 2.119E+03

Wei-Yin Loh 121 GUIDE manual

5.3 Multiple linear: birthwt data 5 REGRESSION

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree same as ++ tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

D-mean is mean of weight in the node

Cases fit give the number of cases used to fit node
MSE and R^2 are based on all cases in node
Node Total Cases Matrix Node Node Node Split Other
label cases fit rank D-mean MSE R^2 variable variables
1 50000 50000 4 3.371E+03 2.989E+05 0.0683 black
2T 8142 8142 4 3.163E+03 3.541E+05 0.0600 boy
3 41858 41858 4 3.411E+03 2.797E+05 0.0673 boy
6T 20229 20229 4 3.352E+03 2.598E+05 0.0679 married
7T 21629 21629 4 3.467E+03 2.923E+05 0.0671 married

Number of terminal nodes of final tree: 3

Total number of nodes of final tree: 5
Second best split variable (based on curvature test) at root node is boy

Regression tree:

Node 1: black = "1"

Node 2: weight-mean = 3.16268E+03
Node 1: black /= "1"
Node 3: boy = "0"
Node 6: weight-mean = 3.35162E+03
Node 3: boy /= "0"
Node 7: weight-mean = 3.46698E+03

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse
variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if black = "1"

Wei-Yin Loh 122 GUIDE manual

5.3 Multiple linear: birthwt data 5 REGRESSION

black mode = "0"

Coefficients of least squares regression function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 2.8271E+03 206.14 0.0000
age 1.0226E+01 23.88 0.0000 1.8000E+01 2.7416E+01 4.5000E+01
cigsper -1.3956E+01 -26.51 0.0000 0.0000E+00 1.4766E+00 6.0000E+01
wtgain 9.2434E+00 48.57 0.0000 0.0000E+00 3.0709E+01 9.8000E+01
Mean of weight = 3370.75664000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 2: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 2.7573E+03 82.48 0.0000
age 5.8561E+00 5.15 0.0000 1.8000E+01 2.5886E+01 4.5000E+01
cigsper -2.0286E+01 -9.79 0.0000 0.0000E+00 8.2031E-01 4.0000E+01
wtgain 9.2828E+00 19.90 0.0000 0.0000E+00 2.9133E+01 9.8000E+01
Mean of weight = 3162.67587816261
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if boy = "0"
boy mode = "1"
----------------------------
Node 6: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 2.9016E+03 140.32 0.0000
age 8.3560E+00 13.14 0.0000 1.8000E+01 2.7715E+01 4.5000E+01
cigsper -1.6363E+01 -21.80 0.0000 0.0000E+00 1.5738E+00 6.0000E+01
wtgain 7.9611E+00 27.98 0.0000 0.0000E+00 3.0667E+01 9.8000E+01
Mean of weight = 3351.62123683820
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 7: Terminal node
Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 2.9362E+03 137.04 0.0000
age 9.5194E+00 14.55 0.0000 1.8000E+01 2.7713E+01 4.5000E+01
cigsper -1.3434E+01 -18.01 0.0000 0.0000E+00 1.6328E+00 6.0000E+01
wtgain 9.2174E+00 31.32 0.0000 0.0000E+00 3.1342E+01 9.8000E+01
Mean of weight = 3466.98317074298
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------

Proportion of variance (R-squared) explained by tree model: .0986

Wei-Yin Loh 123 GUIDE manual

5.3 Multiple linear: birthwt data 5 REGRESSION

black
=1 1

boy
2 =0 3
8142
3163.
6 7
20229 21629
3352. 3467.

Figure 12: GUIDE v.26.0 0.50-SE least-squares multiple linear regression tree for
predicting weight. At each split, an observation goes to the left branch if and only if
the condition is satisfied. Sample sizes (in italics) and means of weight are printed
below nodes. Second best split variable (based on curvature test) at root node is
boy.

Observed and fitted values are stored in mul.fit

Regressor names and coefficients are stored in mul.reg
LaTeX code for tree is in mul.tex
Split and fit variable names are stored in mul.var

Figure 12 shows the piecewise multiple linear model. Although it is slightly

smaller than the piecewise best simple linear model, it may be more difficult o inter-
pret due to the many linear predictor variables in each node.

5.3.3 Contents of mul.var

The contents of mul.var follow.
1 c black black 1 "1"
2 t boy boy 0.3164445923E+04
1 c black black 1 "1"
3 c boy boy 1 "0"
6 t married married 0.3351621237E+04
3 c boy boy 1 "0"
7 t married married 0.3467528370E+04

Wei-Yin Loh 124 GUIDE manual

5.4 Stepwise linear: birthwt data 5 REGRESSION

5.3.4 Contents of mul.reg

The file mul.reg give the node number and the regression coefficients in each node.
Node Constant age cigsper wtgain
2 2.7573E+03 5.8561E+00 -2.0286E+01 9.2828E+00
6 2.9016E+03 8.3560E+00 -1.6363E+01 7.9611E+00
7 2.9362E+03 9.5194E+00 -1.3434E+01 9.2174E+00

5.4 Stepwise linear: birthwt data

Yet another option is to fit a stepwise linear regression model in each node. This
may be better than the piecewise multiple liner model if it reduces the number of
linear predictors in some of the nodes.

5.4.1 Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: step.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: step.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1):
Input 1 for least squares, 2 least median of squares ([1:2], <cr>=1):
Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 0 for stepwise regression in each node (R var. may not be included),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
0: stepwise linear, 1: multiple linear, 2: best polynomial, 3: constant,
4: stepwise simple ANCOVA ([0:4], <cr>=0):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):2
Input 1 for forward+backward, 2 for forward, 3 for all subsets ([1:3], <cr>=1):

Wei-Yin Loh 125 GUIDE manual

5.4 Stepwise linear: birthwt data 5 REGRESSION

Input the maximum number of variables to be selected

0 indicates that the largest possible value is used
Input maximum number of variables to be selected ([0:], <cr>=0):
Input F-to-enter value ([0.01:], <cr>=4.00):
Input F-to-delete value ([0.01:], <cr>=3.99):
Choose a truncation method for predicted values:
0: none, 1: node range, 2: +10% node range, 3: global range
Input 0, 1, 2, or 3 ([0:3], <cr>=3):
Input 1 for interaction tests, 2 to skip them ([1:2], <cr>=1):
Input 1 to prune by CV, 2 for no pruning ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: birthwt.dsc
Reading data description file ...
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Reading data file ...
Number of records in data file: 50000
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 50000
Col. no. Categorical variable #levels #missing values
2 black 2 0
3 married 2 0
4 boy 2 0
6 smoke 2 0
9 visit 4 0
10 ed 4 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished processing 5000 of 50000 observations
Finished processing 10000 of 50000 observations
Finished processing 15000 of 50000 observations
Finished processing 20000 of 50000 observations
Finished processing 25000 of 50000 observations
Finished processing 30000 of 50000 observations
Finished processing 35000 of 50000 observations
Finished processing 40000 of 50000 observations
Finished processing 45000 of 50000 observations
Finished processing 50000 of 50000 observations
Data checks complete
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var

Wei-Yin Loh 126 GUIDE manual

5.4 Stepwise linear: birthwt data 5 REGRESSION

50000 0 0 1 3 0 0 0 6
No weight variable in data file
No. cases used for training: 50000
Finished reading data file
Choose how you wish to deal with missing values in training or test data:
Option 1: Fit separate models to complete and incomplete cases
Option 2: Impute missing F and N values at each node with means for regression
Option 3: Fit a piecewise constant model
Input selection: ([1:3], <cr>=2):
Default number of cross-validations: 10
Input 1 to accept the default, 2 to change it ([1:2], <cr>=1):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50):
Choose fraction of cases for splitting
Larger values give more splits: 0 = median split and 1 = all possible splits
Default fraction is 1.0000
Choose 1 to accept default split fraction, 2 to change it
Input 1 or 2 ([1:2], <cr>=1):
Default max. number of split levels: 30
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 2499
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): step.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1):
Input 1 to number all nodes, 2 to number leaves only ([1:2], <cr>=1):
Choose a color for the terminal nodes:
(1) white
(2) lightgray
(3) gray
(4) darkgray
(5) black
(6) yellow
(7) red
(8) blue
(9) green
(10) magenta
(11) cyan
Input your choice ([1:11], <cr>=6):
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values
Input your choice ([1:3], <cr>=1):3
Input file name: step.var
Input 2 to save regressor names in a file, 1 otherwise ([1:2], <cr>=1):2

Wei-Yin Loh 127 GUIDE manual

5.5 Best ANCOVA: birthwt data 5 REGRESSION

Input file name: step.reg

Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: step.fit
Input 2 to write R function for predicting new cases, 1 otherwise ([1:2], <cr>=1):
Input file is created!
Run GUIDE with the command: guide < step.in

The result is the same as that for the multiple linear regression option in this
example.

5.4.2 Contents of step.reg

The contents of step.reg are slightly different from that of mul.reg. Instead of
giving the estimated regression coefficients in each node, it gives the names of the
variables selected to fit each node. The node number is given in column 1 and the
lower and upper truncation limits in columns 2 and 3.
node lower upper variables
2 2.400E+02 6.350E+03 age cigsper wtgain
6 2.400E+02 6.350E+03 age cigsper wtgain
7 2.400E+02 6.350E+03 age cigsper wtgain

5.5 Best ANCOVA: birthwt data

In the best simple polynomial model, categorical variables that are specified as C
are used only to split the nodes. Sometimes, it may be desired to let them also
serve as linear predictors by means of their dummy variables. This can be done in
the multiple linear and stepwise linear options by simply specifying the categorical
variables as B instead of C. The same can also be done in the best simple polynomial
model, but this has the undesirable effect that a single dummy variable may be
chosen as the best linear predictor in a node. A better alternative is the best simple
ANCOVA option, where at each node, (i) a single N or F variable is selected as the
best linear predictor and (ii) stepwise regression is used to select a subset of the
dummy variables as additional predictors. We demonstrate this by first editing the
description file so that C variables are changed to B as follows. The resulting file is
named birthwtancova.dsc.
birthwt.dat
NA
1
1 weight d
2 black b
3 married b

Wei-Yin Loh 128 GUIDE manual

5.5 Best ANCOVA: birthwt data 5 REGRESSION

4 boy b
5 age n
6 smoke b
7 cigsper n
8 wtgain n
9 visit b
10 ed b
11 lowbwt x

5.5.1 Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: ancova.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: ancova.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1):
Input 1 for least squares, 2 least median of squares ([1:2], <cr>=1):
Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 0 for stepwise regression in each node (R var. may not be included),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
0: stepwise linear, 1: multiple linear, 2: best polynomial, 3: constant,
4: stepwise simple ANCOVA ([0:4], <cr>=0):4
This is where the ANCOVA option is selected.
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):2
Input the maximum number of variables to be selected
0 indicates that the largest possible value is used
Input maximum number of variables to be selected ([0:], <cr>=0):
Input F-to-enter value ([0.01:], <cr>=4.00):
Input F-to-delete value ([0.01:], <cr>=3.99):
Choose a truncation method for predicted values:
0: none, 1: node range, 2: +10% node range, 3: global range,

Wei-Yin Loh 129 GUIDE manual

5.5 Best ANCOVA: birthwt data 5 REGRESSION

4: 2-sided Winsorization Winsorization

Input 0, 1, 2, 3, or 4 ([0:4], <cr>=3):
Input 1 for interaction tests, 2 to skip them ([1:2], <cr>=1):
Input 1 to prune by CV, 2 for no pruning ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: birthwtancova.dsc
Reading data description file ...
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Reading data file ...
Number of records in data file: 50000
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 50000
Col. no. Categorical variable #levels #missing values
2 black 2 0
3 married 2 0
4 boy 2 0
6 smoke 2 0
9 visit 4 0
10 ed 4 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished processing 5000 of 50000 observations
Finished processing 10000 of 50000 observations
Finished processing 15000 of 50000 observations
Finished processing 20000 of 50000 observations
Finished processing 25000 of 50000 observations
Finished processing 30000 of 50000 observations
Finished processing 35000 of 50000 observations
Finished processing 40000 of 50000 observations
Finished processing 45000 of 50000 observations
Finished processing 50000 of 50000 observations
Data checks complete
GUIDE will try to create the variables in the description file.
If it is unsuccessful, please create the columns yourself...
Number of dummy variables created: 10
Creating dummy variables
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 6 0
No weight variable in data file

Wei-Yin Loh 130 GUIDE manual

5.5 Best ANCOVA: birthwt data 5 REGRESSION

No. cases used for training: 50000

Finished reading data file
Choose how you wish to deal with missing values in training or test data:
Option 1: Fit separate models to complete and incomplete cases
Option 2: Impute missing F and N values at each node with means for regression
Option 3: Fit a piecewise constant model
Input selection: ([1:3], <cr>=2):
Default number of cross-validations: 10
Input 1 to accept the default, 2 to change it ([1:2], <cr>=1):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50):
Choose fraction of cases for splitting
Larger values give more splits: 0 = median split and 1 = all possible splits
Default fraction is 1.0000
Choose 1 to accept default split fraction, 2 to change it
Input 1 or 2 ([1:2], <cr>=1):
Default max. number of split levels: 30
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 2499
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): ancova.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1):
Input 1 to number all nodes, 2 to number leaves only ([1:2], <cr>=1):
Choose color(s) for the terminal nodes:
(1) red-yellow-green
(2) red-green-blue
(3) magenta-yellow-green
(4) yellow
(5) green
(6) magenta
(7) cyan
(8) lightgray
(9) white
Input your choice ([1:9], <cr>=1):
You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values
Input your choice ([1:3], <cr>=1):3
Input file name: ancova.var
Input 2 to save truncation limits and regression coefficients in a file, 1 otherwise ([1:2], <cr>=1
Input file name: ancova.reg
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: ancova.fit
Input 2 to write R function for predicting new cases, 1 otherwise ([1:2], <cr>=1):

Wei-Yin Loh 131 GUIDE manual

5.5 Best ANCOVA: birthwt data 5 REGRESSION

Input file is created!

Run GUIDE with the command: guide < ancova.in

5.5.2 Results
Least squares regression tree
Predictions truncated at global min. and max. of D sample values
Pruning by cross-validation
Data description file: birthwtancova.dsc
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Piecewise simple linear ANCOVA model
F-to-enter and F-to-delete: 4.000 3.990
Number of records in data file: 50000
Length of longest data entry: 4
Number of dummy variables created: 10

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 weight d 2.4000E+02 6.3500E+03
2 black b 2
3 married b 2
4 boy b 2
5 age n 1.8000E+01 4.5000E+01
6 smoke b 2
7 cigsper n 0.0000E+00 6.0000E+01
8 wtgain n 0.0000E+00 9.8000E+01
9 visit b 4
10 ed b 4
=================== Constructed variables ===================
12 black.1 f 0.0000E+00 1.0000E+00
13 marri.1 f 0.0000E+00 1.0000E+00
14 boy.1 f 0.0000E+00 1.0000E+00
15 smoke.1 f 0.0000E+00 1.0000E+00
16 visit.1 f 0.0000E+00 1.0000E+00
17 visit.2 f 0.0000E+00 1.0000E+00
18 visit.3 f 0.0000E+00 1.0000E+00
19 ed.1 f 0.0000E+00 1.0000E+00
20 ed.2 f 0.0000E+00 1.0000E+00
21 ed.3 f 0.0000E+00 1.0000E+00

Wei-Yin Loh 132 GUIDE manual

5.5 Best ANCOVA: birthwt data 5 REGRESSION

Indicator F variables are created for the B variables,

with the alphabetically first category of each variable set as reference level.

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 6 0
No weight variable in data file
No. cases used for training: 50000

Missing values imputed with node means for regression

Interaction tests on all variables
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node: 1.0000
Max. number of split levels: 30
Min. node sample size: 2499
Number of SEs for pruned tree: 5.0000E-01

Size and CV MSE and SE of subtrees:

Tree #Tnodes Mean MSE SE(Mean) BSE(Mean) Median MSE BSE(Median)
1 14 2.869E+05 2.778E+03 1.766E+03 2.853E+05 2.102E+03
2 13 2.868E+05 2.777E+03 1.780E+03 2.852E+05 2.106E+03
3 12 2.867E+05 2.775E+03 1.773E+03 2.851E+05 2.073E+03
4 11 2.867E+05 2.773E+03 1.773E+03 2.851E+05 1.989E+03
5 10 2.866E+05 2.772E+03 1.744E+03 2.851E+05 1.979E+03
6+ 9 2.866E+05 2.772E+03 1.753E+03 2.848E+05 2.018E+03
7 8 2.866E+05 2.772E+03 1.754E+03 2.850E+05 1.988E+03
8 7 2.866E+05 2.771E+03 1.813E+03 2.851E+05 1.942E+03
9 5 2.867E+05 2.769E+03 1.861E+03 2.852E+05 2.030E+03
10 4 2.866E+05 2.768E+03 1.909E+03 2.852E+05 2.032E+03
11++ 3 2.865E+05 2.765E+03 1.838E+03 2.855E+05 2.045E+03
12 2 2.867E+05 2.763E+03 1.936E+03 2.861E+05 2.370E+03
13** 1 2.873E+05 2.766E+03 1.904E+03 2.868E+05 2.273E+03

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree same as -- tree
* tree same as ++ tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Wei-Yin Loh 133 GUIDE manual

5.5 Best ANCOVA: birthwt data 5 REGRESSION

D-mean is mean of weight in the node

Cases fit give the number of cases used to fit node
MSE and R^2 are based on all cases in node
Node Total Cases Matrix Node Node Node Split Other
label cases fit rank D-mean MSE R^2 variable variables
1T 50000 50000 9 3.371E+03 2.873E+05 0.1047 age +wtgain
Best split at root node is on age at 20.5000000000000

Number of terminal nodes of final tree: 1

Total number of nodes of final tree: 1
Second best split variable (based on curvature test) at root node is wtgain

Regression tree:

Node 1: weight-mean = 3.37076E+03

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse
variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Terminal node

Coefficients of least squares regression functions:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 3.0354E+03 336.72 0.0000
wtgain 8.5430E+00 45.72 0.0000 0.0000E+00 3.0709E+01 9.8000E+01
black.1 -1.9397E+02 -27.63 0.0000 0.0000E+00 1.6284E-01 1.0000E+00
boy.1 1.0980E+02 22.88 0.0000 0.0000E+00 5.1584E-01 1.0000E+00
ed.1 2.3350E+01 3.65 0.0000 0.0000E+00 2.4258E-01 1.0000E+00
ed.2 4.7962E+01 7.26 0.0000 0.0000E+00 2.4898E-01 1.0000E+00
ed.3 -2.9972E+01 -4.08 0.0000 0.0000E+00 1.5946E-01 1.0000E+00
marri.1 8.6933E+01 14.28 0.0000 0.0000E+00 7.1262E-01 1.0000E+00
smoke.1 -2.0540E+02 -27.71 0.0000 0.0000E+00 1.3066E-01 1.0000E+00
Mean of weight = 3370.75664000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------

Proportion of variance (R-squared) explained by tree model: .1047

Observed and fitted values are stored in ancova.fit

Regressor names and coefficients are stored in ancova.reg
LaTeX code for tree is in ancova.tex
Split and fit variable names are stored in ancova.var

Wei-Yin Loh 134 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

The results show that the tree is trivial, with no splits after pruning. The model is
linear in wtgain and all indicator variables except those for visit.

5.5.3 Contents of ancova.reg

This file gives the estimated regression coefficients in each node. Variables not in-
cluded in the regression have value 0.
node selected lower upper constant age cigsper wtgain black.1 boy.1
1 wtgain 2.4000E+02 6.3500E+03 3.0354E+03 0.0000E+00 0.0000E+00 8.5430E+00 -1.9397E+02 1.0980E+

5.6 Quantile regression: birthwt data

Low birthweight is a term used to describe babies who are born weighing less than
2,500 grams (5 pounds, 8 ounces). In contrast, the average newborn weighs about
8 pounds. Over 8 percent of all newborn babies in the United States have low birth-
weight. We can use GUIDE to estimate conditional 0.08 quantiles (Chaudhuri and Loh,
2002; Koenker and Bassett, 1978) for the birthwt data.

5.6.1 Piecewise constant: 1 quantile

First we fit a 0.08-quantile piecewise constant model.

5.6.2 Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: q08con.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: q08con.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).

Wei-Yin Loh 135 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

Input choice ([1:6], <cr>=1):2

Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
1: multiple linear, 2: best polynomial, 3: constant ([1:3], <cr>=1):3
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):
Input 1 for 1 quantile, 2 for 2 quantiles ([1:2], <cr>=1):
We fit two quantiles in the next section.
Input quantile probability ([0.00:1.00], <cr>=0.50):0.08

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: birthwt.dsc
Reading data description file ...
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is weight
Reading data file ...
Number of records in data file: 50000
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 50000
Col. no. Categorical variable #levels #missing values
2 black 2 0
3 married 2 0
4 boy 2 0
6 smoke 2 0
9 visit 4 0
10 ed 4 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished processing 5000 of 50000 observations
Finished processing 10000 of 50000 observations
Finished processing 15000 of 50000 observations
Finished processing 20000 of 50000 observations
Finished processing 25000 of 50000 observations
Finished processing 30000 of 50000 observations
Finished processing 35000 of 50000 observations
Finished processing 40000 of 50000 observations
Finished processing 45000 of 50000 observations
Finished processing 50000 of 50000 observations
Data checks complete
Rereading data

Wei-Yin Loh 136 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 0 0 3 0 6
No. cases used for training: 50000
Finished reading data file
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): q08con.tex
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: q08con.fit
Input file is created!
Run GUIDE with the command: guide < q08con.in

5.6.3 Results
Quantile regression tree with quantile probability .0800
Pruning by cross-validation
Data description file: birthwt.dsc
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is weight
Piecewise constant model
Number of records in data file: 50000
Length of longest data entry: 4

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 weight d 2.4000E+02 6.3500E+03
2 black c 2
3 married c 2
4 boy c 2
5 age s 1.8000E+01 4.5000E+01
6 smoke c 2
7 cigsper s 0.0000E+00 6.0000E+01
8 wtgain s 0.0000E+00 9.8000E+01
9 visit c 4
10 ed c 4

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 0 0 3 0 6

Wei-Yin Loh 137 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

No. cases used for training: 50000

Interaction tests on all variables

Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node: 1.0000
Max. number of split levels: 30
Min. node sample size: 250
Number of SEs for pruned tree: 5.0000E-01

Size and CV Loss and SE of subtrees:

Tree #Tnodes Mean Loss SE(Mean) BSE(Mean) Median Loss BSE(Median)
1 132 9.067E+01 6.759E-01 5.032E-01 9.091E+01 4.821E-01
2 131 9.067E+01 6.759E-01 5.032E-01 9.090E+01 4.809E-01
3 130 9.067E+01 6.759E-01 5.031E-01 9.091E+01 4.806E-01
4 129 9.067E+01 6.759E-01 5.035E-01 9.091E+01 4.811E-01
5 128 9.067E+01 6.759E-01 5.035E-01 9.091E+01 4.806E-01
6 127 9.067E+01 6.759E-01 5.031E-01 9.093E+01 4.769E-01
7 126 9.067E+01 6.759E-01 5.032E-01 9.092E+01 4.776E-01
8 125 9.067E+01 6.759E-01 5.026E-01 9.092E+01 4.769E-01
9 124 9.067E+01 6.759E-01 5.024E-01 9.092E+01 4.777E-01
10 123 9.067E+01 6.759E-01 5.024E-01 9.092E+01 4.777E-01
11 122 9.067E+01 6.759E-01 5.024E-01 9.092E+01 4.775E-01
12 121 9.067E+01 6.759E-01 5.024E-01 9.092E+01 4.775E-01
13 120 9.067E+01 6.759E-01 5.024E-01 9.092E+01 4.775E-01
14 119 9.067E+01 6.759E-01 5.024E-01 9.092E+01 4.775E-01
15 118 9.067E+01 6.759E-01 5.024E-01 9.092E+01 4.775E-01
16 116 9.067E+01 6.759E-01 5.022E-01 9.092E+01 4.759E-01
17 114 9.068E+01 6.759E-01 5.023E-01 9.092E+01 4.752E-01
18 112 9.068E+01 6.759E-01 5.023E-01 9.092E+01 4.753E-01
19 111 9.067E+01 6.759E-01 5.023E-01 9.092E+01 4.755E-01
20 110 9.067E+01 6.760E-01 5.010E-01 9.092E+01 4.705E-01
21 109 9.067E+01 6.760E-01 5.015E-01 9.092E+01 4.730E-01
22 108 9.067E+01 6.760E-01 5.024E-01 9.092E+01 4.731E-01
23 107 9.067E+01 6.760E-01 5.035E-01 9.092E+01 4.745E-01
24 106 9.066E+01 6.759E-01 5.084E-01 9.092E+01 4.752E-01
25 103 9.068E+01 6.761E-01 5.096E-01 9.098E+01 4.766E-01
26 102 9.067E+01 6.761E-01 5.102E-01 9.098E+01 4.775E-01
27 101 9.067E+01 6.760E-01 5.135E-01 9.098E+01 4.777E-01
28 99 9.067E+01 6.760E-01 5.131E-01 9.098E+01 4.776E-01
29 98 9.068E+01 6.761E-01 5.132E-01 9.101E+01 4.770E-01
30 96 9.067E+01 6.762E-01 5.137E-01 9.098E+01 4.787E-01
31 95 9.068E+01 6.762E-01 5.126E-01 9.098E+01 4.659E-01
32 94 9.067E+01 6.760E-01 5.131E-01 9.095E+01 4.596E-01
33 93 9.067E+01 6.762E-01 5.137E-01 9.095E+01 4.639E-01
34 92 9.066E+01 6.761E-01 5.141E-01 9.095E+01 4.664E-01

Wei-Yin Loh 138 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

35 90 9.066E+01 6.761E-01 5.141E-01 9.095E+01 4.664E-01

36 89 9.066E+01 6.762E-01 5.196E-01 9.095E+01 4.870E-01
37 87 9.064E+01 6.762E-01 5.149E-01 9.093E+01 4.838E-01
38 86 9.064E+01 6.760E-01 5.158E-01 9.094E+01 4.813E-01
39 84 9.064E+01 6.760E-01 5.154E-01 9.094E+01 4.810E-01
40 83 9.063E+01 6.760E-01 5.193E-01 9.094E+01 4.862E-01
41 82 9.064E+01 6.759E-01 5.207E-01 9.094E+01 4.872E-01
42 81 9.064E+01 6.759E-01 5.200E-01 9.094E+01 4.876E-01
43 80 9.064E+01 6.759E-01 5.200E-01 9.094E+01 4.876E-01
44 79 9.064E+01 6.759E-01 5.200E-01 9.094E+01 4.876E-01
45 77 9.064E+01 6.759E-01 5.200E-01 9.094E+01 4.876E-01
46 75 9.064E+01 6.759E-01 5.203E-01 9.094E+01 4.897E-01
47 74 9.062E+01 6.755E-01 5.212E-01 9.092E+01 4.878E-01
48 73 9.061E+01 6.755E-01 5.264E-01 9.096E+01 5.082E-01
49 71 9.061E+01 6.756E-01 5.252E-01 9.096E+01 5.106E-01
50 70 9.061E+01 6.757E-01 5.256E-01 9.096E+01 5.114E-01
51 69 9.062E+01 6.756E-01 5.232E-01 9.096E+01 5.047E-01
52 65 9.060E+01 6.757E-01 5.302E-01 9.092E+01 4.956E-01
53 63 9.061E+01 6.758E-01 5.303E-01 9.092E+01 4.966E-01
54 60 9.061E+01 6.758E-01 5.303E-01 9.092E+01 4.963E-01
55 59 9.061E+01 6.759E-01 5.282E-01 9.092E+01 4.906E-01
56 55 9.059E+01 6.759E-01 5.279E-01 9.087E+01 4.941E-01
57 54 9.057E+01 6.762E-01 5.335E-01 9.091E+01 4.850E-01
58 53 9.057E+01 6.762E-01 5.336E-01 9.091E+01 4.856E-01
59 52 9.055E+01 6.761E-01 5.381E-01 9.091E+01 4.919E-01
60 50 9.055E+01 6.760E-01 5.367E-01 9.091E+01 4.894E-01
61 49 9.054E+01 6.759E-01 5.326E-01 9.093E+01 4.889E-01
62 48 9.054E+01 6.759E-01 5.326E-01 9.093E+01 4.889E-01
63 47 9.054E+01 6.759E-01 5.316E-01 9.093E+01 4.889E-01
64 46 9.053E+01 6.759E-01 5.144E-01 9.096E+01 5.013E-01
65 45 9.050E+01 6.757E-01 5.258E-01 9.096E+01 5.125E-01
66 44 9.050E+01 6.756E-01 5.272E-01 9.096E+01 5.192E-01
67 42 9.050E+01 6.760E-01 5.281E-01 9.096E+01 5.255E-01
68 41 9.049E+01 6.757E-01 5.272E-01 9.088E+01 5.296E-01
69 39 9.049E+01 6.757E-01 5.272E-01 9.088E+01 5.296E-01
70 37 9.046E+01 6.756E-01 5.239E-01 9.080E+01 5.014E-01
71 35 9.044E+01 6.750E-01 5.213E-01 9.075E+01 4.811E-01
72 34 9.046E+01 6.751E-01 5.321E-01 9.076E+01 5.015E-01
73 32 9.044E+01 6.746E-01 5.293E-01 9.073E+01 5.009E-01
74 31 9.044E+01 6.746E-01 5.294E-01 9.074E+01 4.988E-01
75 29 9.044E+01 6.747E-01 5.285E-01 9.074E+01 4.988E-01
76 28 9.046E+01 6.751E-01 5.234E-01 9.079E+01 4.919E-01
77 26 9.045E+01 6.751E-01 5.214E-01 9.078E+01 4.938E-01
78 24 9.040E+01 6.754E-01 5.000E-01 9.077E+01 4.495E-01
79 23 9.043E+01 6.750E-01 5.029E-01 9.084E+01 4.580E-01
80 22 9.042E+01 6.748E-01 5.063E-01 9.084E+01 4.648E-01

Wei-Yin Loh 139 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

81 21 9.038E+01 6.749E-01 4.877E-01 9.067E+01 4.048E-01

82 20 9.034E+01 6.746E-01 4.986E-01 9.067E+01 4.281E-01
83 19 9.035E+01 6.747E-01 4.982E-01 9.067E+01 4.228E-01
84 18 9.033E+01 6.747E-01 4.987E-01 9.063E+01 4.202E-01
85 17 9.035E+01 6.750E-01 5.148E-01 9.063E+01 4.213E-01
86 16 9.035E+01 6.750E-01 5.148E-01 9.063E+01 4.213E-01
87 14 9.033E+01 6.749E-01 5.193E-01 9.056E+01 4.311E-01
88* 13 9.030E+01 6.749E-01 5.213E-01 9.052E+01 4.404E-01
89++ 12 9.033E+01 6.756E-01 5.381E-01 9.052E+01 4.416E-01
90 11 9.047E+01 6.778E-01 5.376E-01 9.074E+01 3.997E-01
91-- 10 9.056E+01 6.800E-01 5.416E-01 9.079E+01 3.948E-01
92** 9 9.056E+01 6.798E-01 5.327E-01 9.074E+01 4.035E-01
93 8 9.066E+01 6.793E-01 5.599E-01 9.089E+01 4.685E-01
94 7 9.068E+01 6.803E-01 5.310E-01 9.088E+01 4.128E-01
95 6 9.068E+01 6.803E-01 5.310E-01 9.088E+01 4.128E-01
96 5 9.068E+01 6.803E-01 5.310E-01 9.088E+01 4.128E-01
97 4 9.190E+01 6.947E-01 5.337E-01 9.222E+01 4.955E-01
98 3 9.209E+01 6.959E-01 4.561E-01 9.222E+01 4.304E-01
99 2 9.323E+01 7.052E-01 5.111E-01 9.346E+01 5.032E-01
100 1 9.572E+01 7.520E-01 5.063E-01 9.610E+01 6.260E-01

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
+ tree same as ++ tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

D-quant is quantile of weight in the node

Cases fit give the number of cases used to fit node
Node Total Cases Matrix Node Split Other
label cases fit rank D-quant variable variables
1 50000 50000 1 2.637E+03 wtgain
2 17934 17934 1 2.438E+03 black
4 3536 3536 1 2.155E+03 wtgain
8T 1107 1107 1 1.786E+03 boy
9T 2429 2429 1 2.268E+03 cigsper
5 14398 14398 1 2.523E+03 cigsper
10 12402 12402 1 2.580E+03 wtgain
20T 1616 1616 1 2.381E+03 visit
21T 10786 10786 1 2.608E+03 married
11T 1996 1996 1 2.211E+03 visit

Wei-Yin Loh 140 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

3 32066 32066 1 2.760E+03 black

6T 4606 4606 1 2.580E+03 cigsper
7 27460 27460 1 2.807E+03 cigsper
14 24053 24053 1 2.835E+03 wtgain
28T 12842 12842 1 2.807E+03 ed
29T 11211 11211 1 2.892E+03 married
15T 3407 3407 1 2.608E+03 wtgain

Number of terminal nodes of final tree: 9

Total number of nodes of final tree: 17
Second best split variable (based on curvature test) at root node is cigsper

Regression tree:

Node 1: wtgain <= 25.50000

Node 2: black = "1"
Node 4: wtgain <= 14.50000
Node 8: weight sample quantile = 1.78600E+03
Node 4: wtgain > 14.50000 or NA
Node 9: weight sample quantile = 2.26800E+03
Node 2: black /= "1"
Node 5: cigsper <= 3.50000 or NA
Node 10: wtgain <= 10.50000
Node 20: weight sample quantile = 2.38100E+03
Node 10: wtgain > 10.50000 or NA
Node 21: weight sample quantile = 2.60800E+03
Node 5: cigsper > 3.50000
Node 11: weight sample quantile = 2.21100E+03
Node 1: wtgain > 25.50000 or NA
Node 3: black = "1"
Node 6: weight sample quantile = 2.58000E+03
Node 3: black /= "1"
Node 7: cigsper <= 1.50000 or NA
Node 14: wtgain <= 35.50000 or NA
Node 28: weight sample quantile = 2.80700E+03
Node 14: wtgain > 35.50000
Node 29: weight sample quantile = 2.89200E+03
Node 7: cigsper > 1.50000
Node 15: weight sample quantile = 2.60800E+03

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse

Wei-Yin Loh 141 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if wtgain <= 2.5500000E+01
wtgain mean = 3.0709E+01
Predicted quantile = 2637.00000000000
----------------------------
Node 2: Intermediate node
A case goes into Node 4 if black = "1"
black mode = "0"
----------------------------
Node 4: Intermediate node
A case goes into Node 8 if wtgain <= 1.4500000E+01
wtgain mean = 1.6786E+01
----------------------------
Node 8: Terminal node
Predicted quantile = 1786.00000000000
----------------------------
Node 9: Terminal node
Predicted quantile = 2268.00000000000
----------------------------
Node 5: Intermediate node
A case goes into Node 10 if cigsper <= 3.5000000E+00 or NA
cigsper mean = 1.9121E+00
----------------------------
Node 10: Intermediate node
A case goes into Node 20 if wtgain <= 1.0500000E+01
wtgain mean = 1.8539E+01
----------------------------
Node 20: Terminal node
Predicted quantile = 2381.00000000000
----------------------------
Node 21: Terminal node
Predicted quantile = 2608.00000000000
----------------------------
Node 11: Terminal node
Predicted quantile = 2211.00000000000
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if black = "1"
black mode = "0"
----------------------------
Node 6: Terminal node
Predicted quantile = 2580.00000000000
----------------------------
Node 7: Intermediate node

Wei-Yin Loh 142 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

A case goes into Node 14 if cigsper <= 1.5000000E+00 or NA

cigsper mean = 1.4429E+00
----------------------------
Node 14: Intermediate node
A case goes into Node 28 if wtgain <= 3.5500000E+01 or NA
wtgain mean = 3.7503E+01
----------------------------
Node 28: Terminal node
Predicted quantile = 2807.00000000000
----------------------------
Node 29: Terminal node
Predicted quantile = 2892.00000000000
----------------------------
Node 15: Terminal node
Predicted quantile = 2608.00000000000
----------------------------

Observed and fitted values are stored in q08con.fit

LaTeX code for tree is in q08con.tex

Figure 13 shows the tree model.

5.6.4 Piecewise constant: 2 quantiles

Now we fit a model to simultaneously predict two quantiles. We demonstrate this
for the 0.08 and 0.12 quantiles.

5.6.5 Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: qcon2.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: qcon2.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).

Wei-Yin Loh 143 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

wtgain
25.50 1

black black
=1 2 =1 3

wtgain cigsper cigsper

14.50 4 3.50 5 6 1.50 7
4606
wtgain 2580. wtgain
8 9 10.50 10 11 35.50 14 15
1107 2429 1996 3407
1786. 2268. 2211. 2608.
20 21 28 29
1616 10786 12842 11211
2381. 2608. 2807. 2892.

Figure 13: GUIDE v.26.0 0.50-SE piecewise constant 0.08-quantile regression tree
for predicting weight. At each split, an observation goes to the left branch if and
only if the condition is satisfied. The symbol stands for or missing. Sample
size (in italics) and 0.08-quantiles of weight printed below nodes. Second best split
variable at root node is cigsper.

Wei-Yin Loh 144 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

Input choice ([1:6], <cr>=1):2

Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
1: multiple linear, 2: best polynomial, 3: constant ([1:3], <cr>=1):3
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):
Input 1 for 1 quantile, 2 for 2 quantiles ([1:2], <cr>=1):2
Choose two quantiles here.
Input 1st quantile probability ([0.00:1.00], <cr>=0.25): 0.08
Input 2nd quantile probability ([0.00:1.00], <cr>=0.75): 0.12
Choose the 0.08 and 0.12 quantiles here.

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: birthwt.dsc
Reading data description file ...
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is weight
Reading data file ...
Number of records in data file: 50000
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 50000
Col. no. Categorical variable #levels #missing values
2 black 2 0
3 married 2 0
4 boy 2 0
6 smoke 2 0
9 visit 4 0
10 ed 4 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished processing 5000 of 50000 observations
Finished processing 10000 of 50000 observations
Finished processing 15000 of 50000 observations
Finished processing 20000 of 50000 observations
Finished processing 25000 of 50000 observations
Finished processing 30000 of 50000 observations
Finished processing 35000 of 50000 observations
Finished processing 40000 of 50000 observations
Finished processing 45000 of 50000 observations
Finished processing 50000 of 50000 observations

Wei-Yin Loh 145 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

Data checks complete

Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 0 0 3 0 6
No. cases used for training: 50000
Finished reading data file
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): q2con.tex
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: qcon2.fit
Input file is created!
Run GUIDE with the command: guide < qcon2.in

5.6.6 Results
Dual-quantile regression tree with .0800 and .1200 quantiles
Pruning by cross-validation
Data description file: birthwt.dsc
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is weight
Piecewise constant model
Number of records in data file: 50000
Length of longest data entry: 4

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 weight d 2.4000E+02 6.3500E+03
2 black c 2
3 married c 2
4 boy c 2
5 age s 1.8000E+01 4.5000E+01
6 smoke c 2
7 cigsper s 0.0000E+00 6.0000E+01
8 wtgain s 0.0000E+00 9.8000E+01
9 visit c 4
10 ed c 4

Total #cases w/ #missing

Wei-Yin Loh 146 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 0 0 3 0 6
No. cases used for training: 50000

Interaction tests on all variables

Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node: 1.0000
Max. number of split levels: 30
Min. node sample size: 250
Number of SEs for pruned tree: 5.0000E-01

Size and CV Loss and SE of subtrees:

Tree #Tnodes Mean Loss SE(Mean) BSE(Mean) Median Loss BSE(Median)
1 137 2.061E+02 1.432E+00 1.017E+00 2.068E+02 8.100E-01
2 136 2.061E+02 1.432E+00 1.017E+00 2.068E+02 8.096E-01
3 135 2.061E+02 1.432E+00 1.017E+00 2.068E+02 8.138E-01
4 132 2.061E+02 1.432E+00 1.017E+00 2.068E+02 8.130E-01
5 131 2.061E+02 1.432E+00 1.017E+00 2.068E+02 8.135E-01
6 130 2.061E+02 1.432E+00 1.011E+00 2.067E+02 8.093E-01
7 129 2.061E+02 1.432E+00 1.013E+00 2.067E+02 8.096E-01
8 128 2.061E+02 1.432E+00 1.013E+00 2.067E+02 8.105E-01
9 127 2.061E+02 1.432E+00 1.012E+00 2.067E+02 8.075E-01
10 126 2.061E+02 1.432E+00 1.012E+00 2.067E+02 8.075E-01
11 125 2.061E+02 1.432E+00 1.012E+00 2.067E+02 8.075E-01
12 123 2.061E+02 1.432E+00 1.013E+00 2.067E+02 8.097E-01
13 121 2.061E+02 1.432E+00 1.013E+00 2.067E+02 8.097E-01
14 120 2.061E+02 1.432E+00 1.014E+00 2.067E+02 8.101E-01
15 119 2.061E+02 1.432E+00 1.014E+00 2.067E+02 8.115E-01
16 118 2.061E+02 1.432E+00 1.015E+00 2.067E+02 8.120E-01
17 117 2.061E+02 1.432E+00 1.014E+00 2.067E+02 8.130E-01
18 116 2.061E+02 1.432E+00 1.014E+00 2.067E+02 8.162E-01
19 115 2.061E+02 1.432E+00 1.016E+00 2.067E+02 8.265E-01
20 114 2.061E+02 1.432E+00 1.014E+00 2.067E+02 8.276E-01
21 113 2.061E+02 1.432E+00 1.013E+00 2.067E+02 8.248E-01
22 112 2.061E+02 1.432E+00 1.012E+00 2.067E+02 8.193E-01
23 111 2.061E+02 1.432E+00 1.012E+00 2.067E+02 8.156E-01
24 110 2.061E+02 1.432E+00 1.014E+00 2.067E+02 8.001E-01
25 109 2.061E+02 1.432E+00 1.014E+00 2.067E+02 7.970E-01
26 108 2.061E+02 1.432E+00 1.012E+00 2.067E+02 7.968E-01
27 107 2.061E+02 1.432E+00 1.013E+00 2.067E+02 7.990E-01
28 106 2.061E+02 1.432E+00 1.013E+00 2.067E+02 8.024E-01
29 105 2.061E+02 1.432E+00 1.013E+00 2.067E+02 8.024E-01
30 104 2.061E+02 1.432E+00 1.013E+00 2.067E+02 8.046E-01
31 102 2.061E+02 1.432E+00 1.011E+00 2.067E+02 7.818E-01
32 101 2.061E+02 1.432E+00 1.018E+00 2.067E+02 7.759E-01

Wei-Yin Loh 147 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

33 99 2.061E+02 1.432E+00 1.021E+00 2.067E+02 7.716E-01

34 97 2.061E+02 1.432E+00 1.022E+00 2.067E+02 7.737E-01
35 96 2.061E+02 1.432E+00 1.022E+00 2.067E+02 7.745E-01
36 93 2.061E+02 1.432E+00 1.020E+00 2.067E+02 7.699E-01
37 90 2.061E+02 1.432E+00 1.016E+00 2.066E+02 7.787E-01
38 89 2.061E+02 1.432E+00 1.017E+00 2.067E+02 7.884E-01
39 88 2.061E+02 1.432E+00 1.016E+00 2.067E+02 7.766E-01
40 87 2.061E+02 1.432E+00 1.020E+00 2.067E+02 7.614E-01
41 86 2.061E+02 1.432E+00 1.022E+00 2.067E+02 7.718E-01
42 84 2.061E+02 1.432E+00 1.024E+00 2.067E+02 7.764E-01
43 83 2.061E+02 1.433E+00 1.014E+00 2.067E+02 7.653E-01
44 82 2.061E+02 1.433E+00 1.014E+00 2.067E+02 7.616E-01
45 81 2.061E+02 1.433E+00 1.014E+00 2.067E+02 7.616E-01
46 80 2.061E+02 1.433E+00 1.010E+00 2.067E+02 7.613E-01
47 79 2.060E+02 1.433E+00 1.011E+00 2.065E+02 7.753E-01
48 78 2.060E+02 1.432E+00 1.013E+00 2.065E+02 8.459E-01
49 76 2.059E+02 1.432E+00 1.018E+00 2.065E+02 8.728E-01
50 75 2.059E+02 1.432E+00 1.017E+00 2.065E+02 8.841E-01
51 74 2.059E+02 1.432E+00 1.017E+00 2.065E+02 8.939E-01
52 73 2.059E+02 1.432E+00 1.020E+00 2.065E+02 8.956E-01
53 70 2.059E+02 1.432E+00 1.021E+00 2.065E+02 8.976E-01
54 69 2.059E+02 1.432E+00 1.020E+00 2.065E+02 8.890E-01
55 68 2.059E+02 1.432E+00 1.020E+00 2.065E+02 8.890E-01
56 67 2.059E+02 1.432E+00 1.020E+00 2.065E+02 8.890E-01
57 64 2.059E+02 1.432E+00 1.018E+00 2.065E+02 9.025E-01
58 62 2.059E+02 1.432E+00 1.015E+00 2.065E+02 8.989E-01
59 60 2.059E+02 1.432E+00 1.017E+00 2.065E+02 8.964E-01
60 56 2.059E+02 1.432E+00 1.015E+00 2.064E+02 8.871E-01
61 55 2.059E+02 1.432E+00 1.009E+00 2.064E+02 8.812E-01
62 53 2.058E+02 1.432E+00 1.005E+00 2.064E+02 8.337E-01
63 52 2.058E+02 1.432E+00 1.008E+00 2.063E+02 8.300E-01
64 51 2.058E+02 1.432E+00 1.010E+00 2.062E+02 8.279E-01
65 50 2.058E+02 1.433E+00 1.023E+00 2.061E+02 8.648E-01
66 49 2.057E+02 1.432E+00 1.030E+00 2.061E+02 9.212E-01
67 48 2.057E+02 1.432E+00 1.028E+00 2.060E+02 9.078E-01
68 45 2.057E+02 1.433E+00 1.029E+00 2.060E+02 9.235E-01
69 44 2.057E+02 1.434E+00 1.035E+00 2.060E+02 9.588E-01
70 43 2.057E+02 1.434E+00 1.054E+00 2.059E+02 9.975E-01
71 40 2.058E+02 1.434E+00 1.056E+00 2.059E+02 9.935E-01
72 38 2.057E+02 1.434E+00 1.062E+00 2.059E+02 1.013E+00
73 35 2.057E+02 1.434E+00 1.064E+00 2.059E+02 1.070E+00
74 33 2.057E+02 1.434E+00 1.076E+00 2.059E+02 1.075E+00
75 31 2.057E+02 1.434E+00 1.084E+00 2.059E+02 1.043E+00
76 29 2.057E+02 1.434E+00 1.093E+00 2.059E+02 1.069E+00
77 28 2.057E+02 1.436E+00 1.088E+00 2.059E+02 1.053E+00
78 27 2.057E+02 1.436E+00 1.083E+00 2.058E+02 1.036E+00

Wei-Yin Loh 148 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

79 26 2.057E+02 1.436E+00 1.080E+00 2.059E+02 1.028E+00

80 25 2.056E+02 1.435E+00 1.082E+00 2.057E+02 1.073E+00
81* 24 2.056E+02 1.435E+00 1.098E+00 2.057E+02 1.059E+00
82 23 2.057E+02 1.435E+00 1.098E+00 2.058E+02 1.032E+00
83 22 2.057E+02 1.435E+00 1.075E+00 2.058E+02 1.001E+00
84 21 2.057E+02 1.435E+00 1.073E+00 2.058E+02 9.865E-01
85 19 2.057E+02 1.434E+00 1.046E+00 2.057E+02 9.082E-01
86 18 2.057E+02 1.433E+00 1.035E+00 2.058E+02 8.778E-01
87 17 2.058E+02 1.435E+00 1.022E+00 2.057E+02 8.492E-01
88 16 2.058E+02 1.438E+00 1.020E+00 2.058E+02 8.271E-01
89 15 2.058E+02 1.438E+00 1.022E+00 2.058E+02 8.401E-01
90 14 2.058E+02 1.438E+00 1.023E+00 2.060E+02 8.486E-01
91 13 2.058E+02 1.437E+00 1.009E+00 2.059E+02 8.290E-01
92++ 12 2.060E+02 1.436E+00 1.034E+00 2.062E+02 7.777E-01
93-- 11 2.061E+02 1.437E+00 1.025E+00 2.062E+02 8.037E-01
94** 10 2.063E+02 1.438E+00 1.031E+00 2.064E+02 8.195E-01
95 9 2.065E+02 1.439E+00 1.029E+00 2.071E+02 8.816E-01
96 8 2.066E+02 1.440E+00 1.040E+00 2.071E+02 8.492E-01
97 7 2.067E+02 1.438E+00 1.061E+00 2.071E+02 8.604E-01
98 6 2.069E+02 1.442E+00 9.665E-01 2.072E+02 7.491E-01
99 5 2.071E+02 1.446E+00 9.008E-01 2.075E+02 7.507E-01
100 4 2.090E+02 1.456E+00 1.123E+00 2.099E+02 1.286E+00
101 3 2.100E+02 1.463E+00 8.880E-01 2.103E+02 8.149E-01
102 2 2.122E+02 1.486E+00 1.002E+00 2.128E+02 9.149E-01
103 1 2.173E+02 1.576E+00 1.015E+00 2.178E+02 1.229E+00

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
* tree same as + tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Cases fit give the number of cases used to fit node

Node Total Cases Matrix Node Split Other
label cases fit rank median variable variables
1 50000 50000 1 2.637E+03 wtgain
2 17934 17934 1 2.438E+03 black
4 14398 14398 1 2.523E+03 cigsper
8 12402 12402 1 2.580E+03 age
16T 1784 1784 1 2.415E+03 2.590E+03 wtgain :visit
17T 10618 10618 1 2.608E+03 2.778E+03 wtgain

Wei-Yin Loh 149 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

9T 1996 1996 1 2.211E+03 2.410E+03 visit

5 3536 3536 1 2.155E+03 wtgain
10T 1107 1107 1 1.786E+03 2.140E+03 wtgain
11T 2429 2429 1 2.268E+03 2.466E+03 cigsper
3 32066 32066 1 2.760E+03 black
6 27460 27460 1 2.807E+03 cigsper
12 24053 24053 1 2.835E+03 wtgain
24T 12842 12842 1 2.807E+03 2.920E+03 ed
25 11211 11211 1 2.892E+03 married
50T 8942 8942 1 2.920E+03 3.060E+03 boy
51T 2269 2269 1 2.785E+03 2.920E+03 visit
13T 3407 3407 1 2.608E+03 2.722E+03 wtgain
7T 4606 4606 1 2.580E+03 2.693E+03 cigsper

Number of terminal nodes of final tree: 10

Total number of nodes of final tree: 19
Second best split variable (based on curvature test) at root node is cigsper

Regression tree:

Node 1: wtgain <= 25.50000

Node 2: black = "0"
Node 4: cigsper <= 3.50000 or NA
Node 8: age <= 21.50000
Node 16: weight sample quantiles = 2.41500E+03, 2.59000E+03
Node 8: age > 21.50000 or NA
Node 17: weight sample quantiles = 2.60800E+03, 2.77800E+03
Node 4: cigsper > 3.50000
Node 9: weight sample quantiles = 2.21100E+03, 2.41000E+03
Node 2: black /= "0"
Node 5: wtgain <= 14.50000
Node 10: weight sample quantiles = 1.78600E+03, 2.14000E+03
Node 5: wtgain > 14.50000 or NA
Node 11: weight sample quantiles = 2.26800E+03, 2.46600E+03
Node 1: wtgain > 25.50000 or NA
Node 3: black = "0"
Node 6: cigsper <= 1.50000 or NA
Node 12: wtgain <= 35.50000 or NA
Node 24: weight sample quantiles = 2.80700E+03, 2.92000E+03
Node 12: wtgain > 35.50000
Node 25: married = "1"
Node 50: weight sample quantiles = 2.92000E+03, 3.06000E+03
Node 25: married /= "1"
Node 51: weight sample quantiles = 2.78500E+03, 2.92000E+03
Node 6: cigsper > 1.50000
Node 13: weight sample quantiles = 2.60800E+03, 2.72200E+03

Wei-Yin Loh 150 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

Node 3: black /= "0"

Node 7: weight sample quantiles = 2.58000E+03, 2.69300E+03

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse
variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if wtgain <= 2.5500000E+01
wtgain mean = 3.0709E+01
Sample lower and upper quantiles and median:
2.6370E+03 2.8000E+03 3.4020E+03
The three numbers above are the 0.08, 0.12 and 0.50 quantiles.
----------------------------
Node 2: Intermediate node
A case goes into Node 4 if black = "0"
black mode = "0"
----------------------------
Node 4: Intermediate node
A case goes into Node 8 if cigsper <= 3.5000000E+00 or NA
cigsper mean = 1.9121E+00
----------------------------
Node 8: Intermediate node
A case goes into Node 16 if age <= 2.1500000E+01
age mean = 2.8155E+01
----------------------------
Node 16: Terminal node
Sample lower and upper quantiles and median:
2.4150E+03 2.5900E+03 3.2040E+03
----------------------------
Node 17: Terminal node
Sample lower and upper quantiles and median:
2.6080E+03 2.7780E+03 3.3900E+03
----------------------------
Node 9: Terminal node
Sample lower and upper quantiles and median:
2.2110E+03 2.4100E+03 3.0900E+03
----------------------------
Node 5: Intermediate node
A case goes into Node 10 if wtgain <= 1.4500000E+01
wtgain mean = 1.6786E+01
----------------------------

Wei-Yin Loh 151 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

Node 10: Terminal node

Sample lower and upper quantiles and median:
1.7860E+03 2.1400E+03 3.0550E+03
----------------------------
Node 11: Terminal node
Sample lower and upper quantiles and median:
2.2680E+03 2.4660E+03 3.1270E+03
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if black = "0"
black mode = "0"
----------------------------
Node 6: Intermediate node
A case goes into Node 12 if cigsper <= 1.5000000E+00 or NA
cigsper mean = 1.4429E+00
----------------------------
Node 12: Intermediate node
A case goes into Node 24 if wtgain <= 3.5500000E+01 or NA
wtgain mean = 3.7503E+01
----------------------------
Node 24: Terminal node
Sample lower and upper quantiles and median:
2.8070E+03 2.9200E+03 3.4590E+03
----------------------------
Node 25: Intermediate node
A case goes into Node 50 if married = "1"
married mode = "1"
----------------------------
Node 50: Terminal node
Sample lower and upper quantiles and median:
2.9200E+03 3.0600E+03 3.6000E+03
----------------------------
Node 51: Terminal node
Sample lower and upper quantiles and median:
2.7850E+03 2.9200E+03 3.4600E+03
----------------------------
Node 13: Terminal node
Sample lower and upper quantiles and median:
2.6080E+03 2.7220E+03 3.2880E+03
----------------------------
Node 7: Terminal node
Sample lower and upper quantiles and median:
2.5800E+03 2.6930E+03 3.2890E+03
----------------------------

Observed and fitted values are stored in qcon2.fit

Wei-Yin Loh 152 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

LaTeX code for tree is in qcon2.tex

Figure 14 shows the tree. The sample size and 0.08 and 0.12 quantiles are printed
below each terminal node.

5.6.7 Piecewise simple linear

Next we fit a piecewise best simple linear 0.08-quantile model.

5.6.8 Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: q08lin.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: q08lin.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1):2
Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
1: multiple linear, 2: best polynomial, 3: constant ([1:3], <cr>=1):2
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):
Input quantile probability ([0.00:1.00], <cr>=0.50):.08

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: birthwt.dsc
Reading data description file ...
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Reading data file ...

Wei-Yin Loh 153 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

wtgain
25.50 1

black black
=0 2 =0 3

cigsper wtgain cigsper

3.50 4 14.50 5 1.50 6 7
4606
2693.
2580.
age wtgain
21.50 8 9 10 11 35.50 12 13
1996 1107 2429 3407
2410. 2140. 2466. 2722.
2211. 1786. 2268. 2608.
married
16 17 24 =1 25
1784 10618 12842
2590. 2778. 2920.
2415. 2608. 2807.

50 51
8942 2269
3060. 2920.
2920. 2785.

Figure 14: GUIDE v.26.0 0.50-SE piecewise constant 0.08 and 0.12-quantile regres-
sion tree for predicting weight. At each split, an observation goes to the left branch
if and only if the condition is satisfied. The symbol stands for or missing.
Sample size (in italics) and sample 0.12 and 0.08-quantiles of weight printed below
nodes. Second best split variable at root node is cigsper.

Wei-Yin Loh 154 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

Number of records in data file: 50000

Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 50000
Col. no. Categorical variable #levels #missing values
2 black 2 0
3 married 2 0
4 boy 2 0
6 smoke 2 0
9 visit 4 0
10 ed 4 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished processing 5000 of 50000 observations
Finished processing 10000 of 50000 observations
Finished processing 15000 of 50000 observations
Finished processing 20000 of 50000 observations
Finished processing 25000 of 50000 observations
Finished processing 30000 of 50000 observations
Finished processing 35000 of 50000 observations
Finished processing 40000 of 50000 observations
Finished processing 45000 of 50000 observations
Finished processing 50000 of 50000 observations
Data checks complete
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 0 6
No. cases used for training: 50000
Finished reading data file
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): q08lin.tex
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: q08lin.fit
Input file is created!
Run GUIDE with the command: guide < q08lin.in

5.6.9 Results
Quantile regression tree with quantile probability .0800
No truncation of predicted values
Pruning by cross-validation
Data description file: birthwt.dsc
Training sample file: birthwt.dat
Missing value code: NA

Wei-Yin Loh 155 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

Records in data file start on line 1

Dependent variable is weight
Piecewise simple linear or constant model
Powers are dropped if they are not significant at level 1.0000
Number of records in data file: 50000
Length of longest data entry: 4

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 weight d 2.4000E+02 6.3500E+03
2 black c 2
3 married c 2
4 boy c 2
5 age n 1.8000E+01 4.5000E+01
6 smoke c 2
7 cigsper n 0.0000E+00 6.0000E+01
8 wtgain n 0.0000E+00 9.8000E+01
9 visit c 4
10 ed c 4

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 0 6
No. cases used for training: 50000

Missing values imputed with node means for regression

Interaction tests on all variables
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node: 1.0000
Max. number of split levels: 30
Min. node sample size: 2499
Number of SEs for pruned tree: 5.0000E-01

Size and CV Loss and SE of subtrees:

Tree #Tnodes Mean Loss SE(Mean) BSE(Mean) Median Loss BSE(Median)
1 15 9.019E+01 6.772E-01 4.993E-01 9.035E+01 4.145E-01
2 14 9.020E+01 6.769E-01 4.977E-01 9.036E+01 4.188E-01
3 13 9.022E+01 6.766E-01 4.975E-01 9.038E+01 4.103E-01
4 12 9.022E+01 6.766E-01 4.975E-01 9.038E+01 4.103E-01
5 11 9.021E+01 6.760E-01 5.026E-01 9.038E+01 4.213E-01
6 10 9.019E+01 6.756E-01 5.050E-01 9.034E+01 4.386E-01
7 9 9.019E+01 6.752E-01 5.026E-01 9.034E+01 4.407E-01

Wei-Yin Loh 156 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

8* 8 9.019E+01 6.752E-01 5.026E-01 9.034E+01 4.407E-01

9 7 9.020E+01 6.753E-01 5.045E-01 9.034E+01 4.445E-01
10+ 6 9.022E+01 6.761E-01 4.922E-01 9.029E+01 4.508E-01
11 5 9.028E+01 6.808E-01 4.854E-01 9.038E+01 3.661E-01
12** 4 9.028E+01 6.808E-01 4.854E-01 9.038E+01 3.661E-01
13 3 9.070E+01 6.876E-01 4.902E-01 9.107E+01 4.263E-01
14 2 9.171E+01 6.875E-01 5.288E-01 9.195E+01 3.940E-01
15 1 9.308E+01 7.061E-01 5.258E-01 9.322E+01 4.314E-01

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree same as ++ tree
** tree same as -- tree
++ tree same as -- tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

D-quant is quantile of weight in the node

Cases fit give the number of cases used to fit node
Node Total Cases Matrix Node Split Other
label cases fit rank D-quant variable variables
1 50000 50000 2 2.637E+03 black
2T 8142 8142 2 2.381E+03 wtgain
3 41858 41858 2 2.710E+03 cigsper
6 36294 36294 2 2.750E+03 married
12T 6561 6561 2 2.637E+03 wtgain
13T 29733 29733 2 2.778E+03 age
7T 5564 5564 2 2.440E+03 boy

Number of terminal nodes of final tree: 4

Total number of nodes of final tree: 7
Second best split variable (based on curvature test) at root node is married

Regression tree:

Node 1: black = "1"

Node 2: weight sample quantile = 2.38100E+03
Node 1: black /= "1"
Node 3: cigsper <= 1.50000 or NA
Node 6: married = "0"
Node 12: weight sample quantile = 2.63700E+03

Wei-Yin Loh 157 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

Node 6: married /= "0"

Node 13: weight sample quantile = 2.77800E+03
Node 3: cigsper > 1.50000
Node 7: weight sample quantile = 2.44000E+03

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse
variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if black = "1"
black mode = "0"
Coefficients of quantile regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 2.3027E+03
wtgain 1.1333E+01 0.0000E+00 3.0709E+01 9.8000E+01
----------------------------
Node 2: Terminal node
Coefficients of quantile regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 1.9779E+03
wtgain 1.3899E+01 0.0000E+00 2.9133E+01 9.8000E+01
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if cigsper <= 1.5000000E+00 or NA
cigsper mean = 1.6043E+00
----------------------------
Node 6: Intermediate node
A case goes into Node 12 if married = "0"
married mode = "1"
----------------------------
Node 12: Terminal node
Coefficients of quantile regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 2.3712E+03
wtgain 9.0400E+00 0.0000E+00 3.1856E+01 9.8000E+01
----------------------------
Node 13: Terminal node
Coefficients of quantile regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 2.4771E+03
wtgain 1.0107E+01 0.0000E+00 3.0948E+01 9.8000E+01

Wei-Yin Loh 158 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

black
=1 1

cigsper
2 1.50 3
8142
2381.
+wtgain
married
=0 6 7
5564
2440.
+wtgain
12 13
6561 29733
2637. 2778.
+wtgain +wtgain

Figure 15: GUIDE v.26.0 0.50-SE piecewise simple linear .080-quantile regression
tree for predicting weight. At each split, an observation goes to the left branch
if and only if the condition is satisfied. The symbol stands for or missing.
Sample size (in italics), 0.08-quantile of weight, and sign and name of best regressor
printed below nodes. Nodes with negative and positive slopes are colored red and
green, respectively. Second best split variable at root node is married.

----------------------------
Node 7: Terminal node
Coefficients of quantile regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 2.0480E+03
wtgain 1.3612E+01 0.0000E+00 3.0389E+01 9.8000E+01
----------------------------

Observed and fitted values are stored in q08lin.fit

LaTeX code for tree is in q08lin.tex

The tree is shown in Figure 15. Piecewise linear quantile regression with two
quantiles simultaneously is not available at the present time.

Wei-Yin Loh 159 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

5.6.10 Piecewise multiple linear

Next we fit a piecewise multiple linear 0.80-quantile model.

5.6.11 Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: q08mul.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: q08mul.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1):2
Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
1: multiple linear, 2: best polynomial, 3: constant ([1:3], <cr>=1):1
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):
Input quantile probability ([0.00:1.00], <cr>=0.50):.08

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: birthwt.dsc
Reading data description file ...
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Reading data file ...
Number of records in data file: 50000
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 50000
Col. no. Categorical variable #levels #missing values
2 black 2 0
3 married 2 0

Wei-Yin Loh 160 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

4 boy 2 0
6 smoke 2 0
9 visit 4 0
10 ed 4 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished processing 5000 of 50000 observations
Finished processing 10000 of 50000 observations
Finished processing 15000 of 50000 observations
Finished processing 20000 of 50000 observations
Finished processing 25000 of 50000 observations
Finished processing 30000 of 50000 observations
Finished processing 35000 of 50000 observations
Finished processing 40000 of 50000 observations
Finished processing 45000 of 50000 observations
Finished processing 50000 of 50000 observations
Data checks complete
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 0 6
No. cases used for training: 50000
Finished reading data file
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): q08mul.tex
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: q08mul.fit
Input file is created!
Run GUIDE with the command: guide < q08mul.in

5.6.12 Results
Quantile regression tree with quantile probability .0800
No truncation of predicted values
Pruning by cross-validation
Data description file: birthwt.dsc
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Piecewise linear model
Number of records in data file: 50000
Length of longest data entry: 4

Summary information (without x variables)

Wei-Yin Loh 161 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

d=dependent, b=split and fit cat variable using 0-1 dummies,

c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 weight d 2.4000E+02 6.3500E+03
2 black c 2
3 married c 2
4 boy c 2
5 age n 1.8000E+01 4.5000E+01
6 smoke c 2
7 cigsper n 0.0000E+00 6.0000E+01
8 wtgain n 0.0000E+00 9.8000E+01
9 visit c 4
10 ed c 4

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 0 6
No. cases used for training: 50000

Missing values imputed with node means for regression

Interaction tests on all variables
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node: 1.0000
Max. number of split levels: 30
Min. node sample size: 2499
100 bootstrap calibration replicates
Scaling for N variables after bootstrap calibration: 1.250
Number of SEs for pruned tree: 5.0000E-01

Size and CV Loss and SE of subtrees:

Tree #Tnodes Mean Loss SE(Mean) BSE(Mean) Median Loss BSE(Median)
1 13 9.010E+01 6.746E-01 4.676E-01 9.065E+01 4.603E-01
2 10 9.010E+01 6.747E-01 4.688E-01 9.065E+01 4.650E-01
3 9 9.012E+01 6.752E-01 4.678E-01 9.068E+01 4.674E-01
4 8 9.012E+01 6.749E-01 4.843E-01 9.064E+01 4.137E-01
5* 7 9.007E+01 6.747E-01 4.932E-01 9.060E+01 4.148E-01
6 6 9.011E+01 6.746E-01 4.941E-01 9.058E+01 3.765E-01
7+ 5 9.013E+01 6.745E-01 4.719E-01 9.055E+01 3.647E-01
8-- 4 9.018E+01 6.756E-01 4.779E-01 9.060E+01 3.893E-01
9** 3 9.034E+01 6.809E-01 5.328E-01 9.060E+01 3.909E-01
10 2 9.059E+01 6.872E-01 5.121E-01 9.085E+01 3.997E-01
11 1 9.213E+01 7.065E-01 5.197E-01 9.224E+01 3.786E-01

0-SE tree based on mean is marked with *

Wei-Yin Loh 162 GUIDE manual

5.6 Quantile regression: birthwt data 5 REGRESSION

0-SE tree based on median is marked with +

Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree same as ++ tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

D-quant is quantile of weight in the node

Cases fit give the number of cases used to fit node
Node Total Cases Matrix Node Split Other
label cases fit rank D-quant variable variables
1 50000 50000 4 2.637E+03 black
2T 8142 8142 4 2.381E+03 wtgain
3 41858 41858 4 2.710E+03 married
6T 9053 9053 4 2.580E+03 boy
7T 32805 32805 4 2.750E+03 wtgain

Number of terminal nodes of final tree: 3

Total number of nodes of final tree: 5
Second best split variable (based on curvature test) at root node is married

Regression tree:

Node 1: black = "1"

Node 2: weight sample quantile = 2.38100E+03
Node 1: black /= "1"
Node 3: married = "0"
Node 6: weight sample quantile = 2.58000E+03
Node 3: married /= "0"
Node 7: weight sample quantile = 2.75000E+03

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse
variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if black = "1"
black mode = "0"
Coefficients of quantile regression function:

Wei-Yin Loh 163 GUIDE manual

5.7 Least median of squares: birthwt data 5 REGRESSION

Regressor Coefficient Minimum Mean Maximum

Constant 2.1765E+03
age 6.0183E+00 1.8000E+01 2.7416E+01 4.5000E+01
cigsper -1.8493E+01 0.0000E+00 1.4766E+00 6.0000E+01
wtgain 1.1288E+01 0.0000E+00 3.0709E+01 9.8000E+01
----------------------------
Node 2: Terminal node
Coefficients of quantile regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 2.0713E+03
age -2.0740E+00 1.8000E+01 2.5886E+01 4.5000E+01
cigsper -3.5349E+01 0.0000E+00 8.2031E-01 4.0000E+01
wtgain 1.3420E+01 0.0000E+00 2.9133E+01 9.8000E+01
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if married = "0"
married mode = "1"
----------------------------
Node 6: Terminal node
Coefficients of quantile regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 2.4254E+03
age -4.0598E+00 1.8000E+01 2.4050E+01 4.5000E+01
cigsper -1.1956E+01 0.0000E+00 3.3095E+00 6.0000E+01
wtgain 9.8152E+00 0.0000E+00 3.1661E+01 9.8000E+01
----------------------------
Node 7: Terminal node
Coefficients of quantile regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 2.3235E+03
age 4.3902E+00 1.8000E+01 2.8725E+01 4.5000E+01
cigsper -2.0623E+01 0.0000E+00 1.1337E+00 6.0000E+01
wtgain 1.0854E+01 0.0000E+00 3.0838E+01 9.8000E+01
----------------------------

Observed and fitted values are stored in q08mul.fit

LaTeX code for tree is in q08mul.tex

Figure 16 shows the tree.

5.7 Least median of squares: birthwt data

Although median regression may be preferred to least-squares regression if there are
large outliers in a data set, an alternative that is even more robust to outliers is least

Wei-Yin Loh 164 GUIDE manual

5.7 Least median of squares: birthwt data 5 REGRESSION

black
=1 1

married
2 =0 3
8142
2381.

6 7
9053 32805
2580. 2750.

Figure 16: GUIDE v.26.0 0.50-SE multiple linear .080-quantile regression tree for
predicting weight. At each split, an observation goes to the left branch if and only
if the condition is satisfied. Sample size (in italics) and 0.08-quantiles of weight
printed below nodes. Second best split variable at root node is married.

median of squares regression (Rousseeuw and Leroy, 1987). GUIDE can construct
tree models using this criterion. We use the birthwt data for illustration. A session
log of the input file generation is below, followed by the results and the LATEX tree
diagram in Figure 17.

0. Read the warranty disclaimer

1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: lms.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: lms.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1):
Input 1 for least squares, 2 least median of squares ([1:2], <cr>=1):2
This is where the option for least median of squares is selected.
Choose complexity of model to use at each node:

Wei-Yin Loh 165 GUIDE manual

5.7 Least median of squares: birthwt data 5 REGRESSION

If R variable present (i.e., subgroup identification),

choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
1: multiple linear, 2: best simple linear, 3: constant ([1:3], <cr>=2):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: birthwt.dsc
Reading data description file ...
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Reading data file ...
Number of records in data file: 50000
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 50000
Col. no. Categorical variable #levels #missing values
2 black 2 0
3 married 2 0
4 boy 2 0
6 smoke 2 0
9 visit 4 0
10 ed 4 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished processing 5000 of 50000 observations
Finished processing 10000 of 50000 observations
Finished processing 15000 of 50000 observations
Finished processing 20000 of 50000 observations
Finished processing 25000 of 50000 observations
Finished processing 30000 of 50000 observations
Finished processing 35000 of 50000 observations
Finished processing 40000 of 50000 observations
Finished processing 45000 of 50000 observations
Finished processing 50000 of 50000 observations
Data checks complete
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 0 6
No weight variable in data file
No. cases used for training: 50000
Finished reading data file

Wei-Yin Loh 166 GUIDE manual

5.7 Least median of squares: birthwt data 5 REGRESSION

Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):

Input file name to store LaTeX code (use .tex as suffix): lms.tex
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: lms.fit
Input file is created!
Run GUIDE with the command: guide < lms.in

5.7.1 Results
Least median of squares regression tree
Predictions truncated at global min. and max. of D sample values
Pruning by cross-validation
Data description file: birthwt.dsc
Training sample file: birthwt.dat
Missing value code: NA
Records in data file start on line 1
Dependent variable is weight
Piecewise simple linear or constant model
Powers are dropped if they are not significant at level 1.0000
Number of records in data file: 50000
Length of longest data entry: 4

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 weight d 2.4000E+02 6.3500E+03
2 black c 2
3 married c 2
4 boy c 2
5 age n 1.8000E+01 4.5000E+01
6 smoke c 2
7 cigsper n 0.0000E+00 6.0000E+01
8 wtgain n 0.0000E+00 9.8000E+01
9 visit c 4
10 ed c 4

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
50000 0 0 1 3 0 0 0 6
No weight variable in data file
No. cases used for training: 50000

Missing values imputed with node means for regression

Wei-Yin Loh 167 GUIDE manual

5.7 Least median of squares: birthwt data 5 REGRESSION

Interaction tests on all variables

Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node: 1.0000
Max. number of split levels: 30
Min. node sample size: 2499
Number of SEs for pruned tree: 5.0000E-01

Size and CV median absolute residual (MAR) and SE of subtrees:

Tree #Tnodes Mean MAR BSE(Mean) Median MAR BSE(Median)
1 13 1.592E+06 2.036E+00 3.182E+02 5.316E+00
2 11 1.591E+06 1.963E+00 3.176E+02 5.074E+00
3-- 9 1.591E+06 1.963E+00 3.176E+02 5.074E+00
4 8 1.591E+06 2.010E+00 3.166E+02 5.033E+00
5++ 6 1.592E+06 2.244E+00 3.164E+02 3.514E+00
6 5 1.605E+06 2.143E+00 3.205E+02 3.418E+00
7 4 1.603E+06 1.999E+00 3.223E+02 3.486E+00
8 3 1.627E+06 2.529E+00 3.260E+02 2.697E+00
9 1 1.640E+06 9.479E-01 3.263E+02 9.920E-01
The selected tree is marked by two dashes --.
0-SE tree based on mean is marked with *
0-SE tree based on median is marked with +
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
+ tree same as ++ tree
* tree same as -- tree

Following tree is based on mean CV with bootstrap SE estimate (--).

Structure of final tree. Each terminal node is marked with a T.

D-mean is mean of weight in the node

Cases fit give the number of cases used to fit node
MAR is median of absolute residuals
Node Total Cases Matrix Node Node Split Other
label cases fit rank D-median MAR variable variables
1 50000 50000 2 3.402E+03 3.260E+02 cigsper +cigsper
2 43467 43467 2 3.430E+03 3.241E+02 black +wtgain
4 7350 7350 2 3.231E+03 3.256E+02 boy +wtgain
8T 3584 3584 1 3.158E+03 3.116E+02 - *Constant*
9T 3766 3766 2 3.289E+03 3.262E+02 - +wtgain
5 36117 36117 1 3.459E+03 3.120E+02 wtgain *Constant*
10T 13965 13965 1 3.374E+03 3.150E+02 boy *Constant*
11 22152 22152 2 3.515E+03 3.048E+02 boy +wtgain
22T 10500 10500 2 3.459E+03 2.912E+02 wtgain +age
23 11652 11652 2 3.580E+03 3.145E+02 age +wtgain

Wei-Yin Loh 168 GUIDE manual

5.7 Least median of squares: birthwt data 5 REGRESSION

46T 4075 4075 2 3.515E+03 3.082E+02 - +wtgain

47 7577 7577 1 3.629E+03 3.115E+02 wtgain *Constant*
94T 4596 4596 1 3.572E+03 3.021E+02 - *Constant*
95T 2981 2981 2 3.686E+03 3.202E+02 - +wtgain
3 6533 6533 2 3.203E+03 3.202E+02 boy +wtgain
6T 3148 3148 1 3.119E+03 2.976E+02 - *Constant*
7T 3385 3385 2 3.260E+03 3.202E+02 - +wtgain

Number of terminal nodes of final tree: 9

Total number of nodes of final tree: 17
Second best split variable (based on curvature test) at root node is wtgain

Regression tree:

Node 1: cigsper <= 0.50000 or NA

Node 2: black = "1"
Node 4: boy = "0"
Node 8: weight-mean = 3.15750E+03
Node 4: boy /= "0"
Node 9: weight-mean = 3.28900E+03
Node 2: black /= "1"
Node 5: wtgain <= 27.50000
Node 10: weight-mean = 3.37400E+03
Node 5: wtgain > 27.50000 or NA
Node 11: boy = "0"
Node 22: weight-mean = 3.45900E+03
Node 11: boy /= "0"
Node 23: age <= 25.50000
Node 46: weight-mean = 3.51500E+03
Node 23: age > 25.50000 or NA
Node 47: wtgain <= 38.50000 or NA
Node 94: weight-mean = 3.57200E+03
Node 47: wtgain > 38.50000
Node 95: weight-mean = 3.68600E+03
Node 1: cigsper > 0.50000
Node 3: boy = "0"
Node 6: weight-mean = 3.11900E+03
Node 3: boy /= "0"
Node 7: weight-mean = 3.26000E+03

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse

Wei-Yin Loh 169 GUIDE manual

5.7 Least median of squares: birthwt data 5 REGRESSION

variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if cigsper <= 5.0000000E-01 or NA
cigsper mean = 1.4766E+00
Coefficients of least median of squares regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 3.4160E+03
cigsper 7.8333E+00 0.00 1.4766 6.0000E+01
Mean of weight = 3402.00000000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 2: Intermediate node
A case goes into Node 4 if black = "1"
black mode = "0"
----------------------------
Node 4: Intermediate node
A case goes into Node 8 if boy = "0"
boy mode = "1"
----------------------------
Node 8: Terminal node
Coefficients of least median of squares regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 3.2035E+03
wtgain 0.0000E+00 0.0028.9068 9.8000E+01
Mean of weight = 3157.50000000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 9: Terminal node
Coefficients of least median of squares regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 3.1610E+03
wtgain 5.6500E+00 0.0029.7464 9.8000E+01
Mean of weight = 3289.00000000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 5: Intermediate node
A case goes into Node 10 if wtgain <= 2.7500000E+01
wtgain mean = 3.1109E+01
----------------------------
Node 10: Terminal node
Coefficients of least median of squares regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 3.3990E+03
age 0.0000E+00 18.0028.2233 4.5000E+01
Mean of weight = 3374.00000000000

Wei-Yin Loh 170 GUIDE manual

5.7 Least median of squares: birthwt data 5 REGRESSION

Predicted values truncated at 240.000000000000 & 6350.00000000000

----------------------------
Node 11: Intermediate node
A case goes into Node 22 if boy = "0"
boy mode = "1"
----------------------------
Node 22: Terminal node
Coefficients of least median of squares regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 3.0386E+03
age 1.4250E+01 18.0027.8578 4.5000E+01
Mean of weight = 3459.00000000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 23: Intermediate node
A case goes into Node 46 if age <= 2.5500000E+01
age mean = 2.7833E+01
----------------------------
Node 46: Terminal node
Coefficients of least median of squares regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 3.2105E+03
wtgain 8.5000E+00 28.0039.8321 9.8000E+01
Mean of weight = 3515.00000000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 47: Intermediate node
A case goes into Node 94 if wtgain <= 3.8500000E+01 or NA
wtgain mean = 3.7770E+01
----------------------------
Node 94: Terminal node
Coefficients of least median of squares regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 3.5820E+03
wtgain 0.0000E+00 28.0032.3040 3.8000E+01
Mean of weight = 3572.00000000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 95: Terminal node
Coefficients of least median of squares regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 3.3170E+03
wtgain 8.0000E+00 39.0046.1962 9.8000E+01
Mean of weight = 3686.00000000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------

Wei-Yin Loh 171 GUIDE manual

5.8 Poisson regression with offset: lung cancer data 5 REGRESSION

Node 3: Intermediate node

A case goes into Node 6 if boy = "0"
boy mode = "1"
----------------------------
Node 6: Terminal node
Coefficients of least median of squares regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 3.1045E+03
cigsper 0.0000E+00 1.0011.1852 6.0000E+01
Mean of weight = 3119.00000000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------
Node 7: Terminal node
Coefficients of least median of squares regression function:
Regressor Coefficient Minimum Mean Maximum
Constant 2.9496E+03
wtgain 1.0078E+01 0.0030.3412 9.8000E+01
Mean of weight = 3260.00000000000
Predicted values truncated at 240.000000000000 & 6350.00000000000
----------------------------

Proportion of deviance explained by tree model: .0735

Observed and fitted values are stored in lms.fit

LaTeX code for tree is in lms.tex

The tree is shown in Figure 17.

5.8 Poisson regression with offset: lung cancer data

We use a data set from an epidemiological study of the effect of public drinking water
on cancer mortality in Missouri (Choi et al., 2005). Our data file lungcancer.txt
gives the number of deaths (deaths) from lung cancer among 115 counties (county)
during the period 19721981 for both sexes (sex) and four age groups (agegp): 45
54, 5564, 6574, and over 75. The description file lungcancer.dsc below lists
the variables together with the county population (pop) and the natural log of pop
(logpop). The latter is specified as z to serve an an offset variable and the former
is excluded (x) from the analysis. For the purpose of illustration, we specify sex as
b to allow its dummy indicator variable to serve as a linear predictor in the node
Poisson models. The contents of lungcancer.dsc are:
lungcancer.txt
NA
1
1 county c

Wei-Yin Loh 172 GUIDE manual

5.8 Poisson regression with offset: lung cancer data 5 REGRESSION

cigsper
.5000 1

black boy
=1 2 =0 3

boy wtgain
=0 4 27.50 5 6 7
3148 3385
3119. 3260.
+wtgain
boy
8 9 10 =0 11
3584 3766 13965
3158. 3289. 3374.
+wtgain age
22 25.50 23
10500
3459.
+age
wtgain
46 38.50 47
4075
3515.
+wtgain
94 95
4596 2981
3572. 3686.
+wtgain

Figure 17: GUIDE v.26.0 0.50-SE piecewise simple linear least-median-of-squares

regression tree for predicting weight. At each split, an observation goes to the left
branch if and only if the condition is satisfied. The symbol stands for or
missing. Sample size (in italics), mean of weight, and sign and name of regressor
variables printed below nodes. Nodes with negative and positive slopes are colored
red and green, respectively. Second best split variable at root node is wtgain.

Wei-Yin Loh 173 GUIDE manual

5.8 Poisson regression with offset: lung cancer data 5 REGRESSION

2 sex b
3 agegp c
4 deaths d
5 pop x
6 logpop z

Our goal is to construct a Poisson regression tree for the gender-specific rate of lung
cancer deaths, where rate is the expected number of deaths in a county divided by
its population size for each gender. That is, letting denote the expected number
of gender-specific deaths in a county, we fit this model in each node of the tree:

log(/pop) = 0 + 1 I(sex = M)

or, equivalently,
log() = 0 + 1 I(sex = M) + logpop.

5.8.1 Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: poi.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: poi.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1): 2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1): 3
Choose Poisson regression here.
Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
1: multiple linear, 2: best polynomial, 3: constant ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

Wei-Yin Loh 174 GUIDE manual

5.8 Poisson regression with offset: lung cancer data 5 REGRESSION

enclose with matching quotes if it has spaces: lungcancer.dsc

Reading data description file ...
Training sample file: lungcancer.txt
Missing value code: NA
Records in data file start on line 1
Dependent variable is deaths
Reading data file ...
Number of records in data file: 920
Length of longest data entry: 8
Checking for missing values ...
Total number of cases: 920
Col. no. Categorical variable #levels #missing values
1 county 115 0
2 sex 2 0
3 agegp 4 0
Re-checking data ...
Assigning codes to categorical and missing values
Data checks complete
Number of cases with positive D values: 869
GUIDE will try to create the variables in the description file.
If it is unsuccessful, please create the columns yourself...
Number of dummy variables created: 1
Creating dummy variables
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
920 0 0 1 0 0 0 1 2
Offset variable in column: 6
No. cases used for training: 920
Finished reading data file
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): poi.tex
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: poi.fit
Input file is created!
Run GUIDE with the command: guide < poi.in

5.8.2 Results
Poisson regression tree
No truncation of predicted values
Pruning by cross-validation
Data description file: lungcancer.dsc
Training sample file: lungcancer.txt
Missing value code: NA

Wei-Yin Loh 175 GUIDE manual

5.8 Poisson regression with offset: lung cancer data 5 REGRESSION

Records in data file start on line 1

Dependent variable is deaths
Piecewise linear model
Number of records in data file: 920
Length of longest data entry: 8
Number of cases with positive D values: 869
Number of dummy variables created: 1

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
z=offset variable
Column Name Minimum Maximum #Categories #Missing
1 county c 115
2 sex b 2
3 agegp c 4
4 deaths d 0.0000E+00 1.0460E+03
6 logpop z 4.8283E+00 1.0964E+01
=================== Constructed variables ===================
7 sex.M f 0.0000E+00 1.0000E+00

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
920 0 0 1 0 0 0 1 2
Offset variable in column 6
No. cases used for training: 920

Missing values imputed with node means for regression

Interaction tests on all variables
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node: 1.0000
Max. number of split levels: 10
Min. node sample size: 45
Number of SEs for pruned tree: 5.0000E-01

Size and CV Loss and SE of subtrees:

Tree #Tnodes Mean Loss SE(Mean) BSE(Mean) Median Loss BSE(Median)
1 15 2.429E+00 2.831E-01 2.423E-01 2.164E+00 2.871E-01
2 14 2.429E+00 2.831E-01 2.423E-01 2.164E+00 2.871E-01
3 13 2.414E+00 2.829E-01 2.386E-01 2.123E+00 2.857E-01
4 12 2.447E+00 3.157E-01 2.806E-01 2.123E+00 2.805E-01
5 11 2.449E+00 3.157E-01 2.802E-01 2.123E+00 2.786E-01
6 10 2.423E+00 3.153E-01 2.857E-01 2.123E+00 3.114E-01
7 8 2.359E+00 3.153E-01 2.925E-01 1.991E+00 2.698E-01

Wei-Yin Loh 176 GUIDE manual

5.8 Poisson regression with offset: lung cancer data 5 REGRESSION

8 7 2.424E+00 3.563E-01 3.029E-01 1.991E+00 3.756E-01

9 4 2.455E+00 3.772E-01 3.185E-01 1.925E+00 4.297E-01
10** 3 2.318E+00 3.595E-01 3.111E-01 1.709E+00 5.118E-01
11 2 4.857E+00 8.176E-01 5.168E-01 4.388E+00 7.094E-01
12 1 9.439E+00 1.406E+00 1.131E+00 9.048E+00 1.683E+00

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree and ++ tree are the same

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Rate is mean of Y/exp(offset)

Cases fit give the number of cases used to fit node
Deviance is mean residual deviance for all cases in node
Node Total Cases Matrix Node Node Split Other
label cases fit rank rate deviance variable variables
1 920 920 2 1.382E-02 9.179E+00 agegp
2T 230 230 2 5.493E-03 1.863E+00 county
3 690 690 2 1.763E-02 4.357E+00 agegp
6T 230 230 2 1.339E-02 3.003E+00 county
7T 460 460 2 2.093E-02 1.802E+00 agegp

Number of terminal nodes of final tree: 3

Total number of nodes of final tree: 5
Second best split variable (based on curvature test) at root node is sex

Regression tree:

Node 1: agegp = "45-54"

Node 2: deaths sample rate = 5.49286E-03
Node 1: agegp /= "45-54"
Node 3: agegp = "55-64"
Node 6: deaths sample rate = 1.33898E-02
Node 3: agegp /= "55-64"
Node 7: deaths sample rate = 2.09327E-02

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap

Wei-Yin Loh 177 GUIDE manual

5.8 Poisson regression with offset: lung cancer data 5 REGRESSION

solution, see Loh et al. (2016), "Identification of subgroups with

differential treatment effects for longitudinal and multiresponse
variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if agegp = "45-54"
agegp mode = "45-54"
Coefficients of loglinear regression function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant -5.1717E+00 -366.86 0.0000
sex.M 1.4370E+00 89.64 0.0000 0.0000E+00 5.0000E-01 1.0000E+00
Node mean for offset variable = 6.7275E+00
----------------------------
Node 2: Terminal node
Coefficients of loglinear regression function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant -5.8337E+00 -161.46 0.0000
sex.M 1.0384E+00 24.44 0.0000 0.0000E+00 5.0000E-01 1.0000E+00
Node mean for offset variable = 6.8567E+00
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if agegp = "55-64"
agegp mode = "55-64"
----------------------------
Node 6: Terminal node
Coefficients of loglinear regression function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant -5.1175E+00 -199.84 0.0000
sex.M 1.2854E+00 43.87 0.0000 0.0000E+00 5.0000E-01 1.0000E+00
Node mean for offset variable = 6.9199E+00
----------------------------
Node 7: Terminal node
Coefficients of loglinear regression function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant -4.9065E+00 -256.88 0.0000
sex.M 1.7137E+00 79.68 0.0000 0.0000E+00 5.0000E-01 1.0000E+00
Node mean for offset variable = 6.5666E+00
----------------------------

Observed and fitted values are stored in poi.fit

LaTeX code for tree is in poi.tex

The results show that the death rate increases with age and that the rate for
males is consistently higher than that for females. The tree diagram is given in
Figure 18.

Wei-Yin Loh 178 GUIDE manual

5.9 Censored response: heart attack data 5 REGRESSION

agegp
=45-54 1

agegp
2 =55-64 3
230
5.493E-03
6 7
230 460
0.01 0.02

Figure 18: GUIDE v.26.0 0.50-SE multiple linear Poisson regression tree for predict-
ing rate of deaths. At each split, an observation goes to the left branch if and only
if the condition is satisfied. Sample size (in italics) and sample rate printed below
nodes. Second best split variable at root node is sex.

5.9 Censored response: heart attack data

GUIDE can fit a piecewise-constant, piecewise-simple linear, or piecewise multiple
linear proportional hazards regression model to censored response data.Using usual
notation, let (x, t) denote the hazard rate at time t for a subject with covariate
vector x. In a proportional hazards model, the hazard rate can be factored as
(x, t) = 0 (t)f (x, ), where 0 (t) is a baseline hazard rate that is independent of
the covariates and f (x, ) is a function of x and some coefficients , independent of
t. The Cox proportional hazards model uses (x, t) = 0 (t) exp( x). GUIDE fits
the more general model
X
(x, t) = 0 (t) I(x Si ) exp( i x),
i

where Si is a set corresponding node i and i is its associated coefficient vector. See
Loh et al. (2015) for more details. P
We illustrate the piecewise-constant model (x, t) = 0 (t) i I(x Si ) exp(i0 )
with a data set from the Worcester Heart Attack Study analyzed in Hosmer et al.
(2008). The data are in the file whas500.csv and the description file in whas500.dsc
whose contents are repeated below.
whas500.csv
NA
1
1 id x

Wei-Yin Loh 179 GUIDE manual

5.9 Censored response: heart attack data 5 REGRESSION

2 age n
3 gender c
4 hr n
5 sysbp n
6 diasbp n
7 bmi n
8 cvd c
9 afb c
10 sho c
11 chf c
12 av3 c
13 miord c
14 mitype c
15 year c
16 admitdate x
17 disdate x
18 fdate x
19 los n
20 dstat x
21 lenfol t
22 fstat d

The goal of the study is to observe survival rates following hospital admission for
acute myocardial infarction. The response variable is lenfol, which stands for total
length of follow-up in days. Variable fstat is status at last follow-up (0=alive,
1=dead) and variable chf is congestive heart complications (0=no, 1=yes).
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: cons.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: cons.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1): 2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1): 4
Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 1 for multiple regression (including R var.) in each node,

Wei-Yin Loh 180 GUIDE manual

5.9 Censored response: heart attack data 5 REGRESSION

choose 2 to fit one linear prognostic var. (N or F) in each node,

choose 3 (constant) to fit only treatment effect in each node
1: multiple linear, 2: best simple linear, 3: constant ([1:3], <cr>=3):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: whas500.dsc
Reading data description file ...
Training sample file: whas500.csv
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is fstat
Reading data file ...
Number of records in data file: 500
Length of longest data entry: 10
Checking for missing values ...
Total number of cases: 500
Col. no. Categorical variable #levels #missing values
3 gender 2 0
8 cvd 2 0
9 afb 2 0
10 sho 2 0
11 chf 2 0
12 av3 2 0
13 miord 2 0
14 mitype 2 0
15 year 3 0
Re-checking data ...
Assigning codes to categorical and missing values
Data checks complete
Smallest uncensored T: 1.0000
No. complete cases excluding censored T < smallest uncensored T: 500
No. cases used to compute baseline hazard: 500
No. cases with D=1 and T >= smallest uncensored: 215
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
500 0 0 5 0 0 6 0 9
Survival time variable in column: 21
Event indicator variable in column: 22
Proportion uncensored among nonmissing T and D variables: .430
No. cases used for training: 500
Finished reading data file
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): cons.tex

Wei-Yin Loh 181 GUIDE manual

5.9 Censored response: heart attack data 5 REGRESSION

Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: cons.fit
Input file is created!
Run GUIDE with the command: guide < cons.in

5.9.1 Results
Proportional hazards regression with relative risk estimates
Pruning by cross-validation
Data description file: whas500.dsc
Training sample file: whas500.csv
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is fstat
Piecewise constant model
Number of records in data file: 500
Length of longest data entry: 10
Smallest uncensored T: 1.0000
No. complete cases excluding censored T < smallest uncensored T: 500
No. cases used to compute baseline hazard: 500
No. cases with D=1 and T >= smallest uncensored: 215

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
t=survival time variable
Column Name Minimum Maximum #Categories #Missing
2 age s 3.0000E+01 1.0400E+02
3 gender c 2
4 hr s 3.5000E+01 1.8600E+02
5 sysbp s 5.7000E+01 2.4400E+02
6 diasbp s 6.0000E+00 1.9800E+02
7 bmi s 1.3045E+01 4.4839E+01
8 cvd c 2
9 afb c 2
10 sho c 2
11 chf c 2
12 av3 c 2
13 miord c 2
14 mitype c 2
15 year c 3
19 los s 0.0000E+00 4.7000E+01
21 lenfol t 1.0000E+00 2.3580E+03
22 fstat d 0.0000E+00 1.0000E+00

Wei-Yin Loh 182 GUIDE manual

5.9 Censored response: heart attack data 5 REGRESSION

=================== Constructed variables ===================

23 lnbasehaz z -4.1352E+00 9.7549E-01

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
500 0 0 5 0 0 6 0 9
Survival time variable in column: 21
Event indicator variable in column: 22
Proportion uncensored among nonmissing T and D variables: 0.430
No. cases used for training: 500

Interaction tests on all variables

Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node: 1.0000
Max. number of split levels: 10
Min. node sample size: 2
Number of iterations: 5
Number of SEs for pruned tree: 5.0000E-01

Size and CV Loss and SE of subtrees:

Tree #Tnodes Mean Loss SE(Mean) BSE(Mean) Median Loss BSE(Median)
1 77 1.783E+00 1.248E-01 7.561E-02 1.803E+00 9.327E-02
2 76 1.783E+00 1.248E-01 7.561E-02 1.803E+00 9.327E-02
3 75 1.782E+00 1.248E-01 7.524E-02 1.800E+00 9.293E-02
4 74 1.784E+00 1.248E-01 7.619E-02 1.800E+00 9.299E-02
5 73 1.783E+00 1.248E-01 7.597E-02 1.800E+00 9.276E-02
6 72 1.783E+00 1.250E-01 7.660E-02 1.800E+00 9.698E-02
7 71 1.783E+00 1.250E-01 7.660E-02 1.800E+00 9.698E-02
8 70 1.783E+00 1.250E-01 7.660E-02 1.800E+00 9.698E-02
9 69 1.783E+00 1.250E-01 7.665E-02 1.800E+00 9.753E-02
10 67 1.786E+00 1.252E-01 7.681E-02 1.816E+00 9.644E-02
11 66 1.775E+00 1.239E-01 7.287E-02 1.816E+00 9.553E-02
12 65 1.769E+00 1.238E-01 7.343E-02 1.816E+00 1.044E-01
13 64 1.770E+00 1.245E-01 7.420E-02 1.834E+00 1.122E-01
14 63 1.764E+00 1.242E-01 7.089E-02 1.828E+00 1.056E-01
15 62 1.764E+00 1.243E-01 7.400E-02 1.828E+00 1.070E-01
16 61 1.756E+00 1.219E-01 7.634E-02 1.828E+00 1.193E-01
17 60 1.748E+00 1.212E-01 7.894E-02 1.828E+00 1.204E-01
18 59 1.714E+00 1.181E-01 6.570E-02 1.789E+00 1.198E-01
19 57 1.714E+00 1.181E-01 6.570E-02 1.789E+00 1.198E-01
20 54 1.712E+00 1.181E-01 6.511E-02 1.789E+00 1.185E-01
21 51 1.720E+00 1.198E-01 6.828E-02 1.789E+00 1.214E-01
22 50 1.717E+00 1.198E-01 6.693E-02 1.789E+00 1.206E-01
23 39 1.714E+00 1.198E-01 6.722E-02 1.789E+00 1.257E-01
24 36 1.713E+00 1.198E-01 6.686E-02 1.789E+00 1.255E-01

Wei-Yin Loh 183 GUIDE manual

5.9 Censored response: heart attack data 5 REGRESSION

25 35 1.702E+00 1.177E-01 6.532E-02 1.769E+00 1.110E-01

26 34 1.703E+00 1.179E-01 6.531E-02 1.769E+00 1.110E-01
27 33 1.696E+00 1.177E-01 6.785E-02 1.769E+00 1.194E-01
28 32 1.701E+00 1.180E-01 6.770E-02 1.769E+00 1.129E-01
29 31 1.679E+00 1.152E-01 6.351E-02 1.698E+00 1.048E-01
30 30 1.680E+00 1.152E-01 6.335E-02 1.698E+00 1.048E-01
31 28 1.680E+00 1.152E-01 6.335E-02 1.698E+00 1.048E-01
32 25 1.660E+00 1.144E-01 5.910E-02 1.681E+00 9.006E-02
33 24 1.642E+00 1.131E-01 5.746E-02 1.652E+00 9.040E-02
34 22 1.620E+00 1.103E-01 5.354E-02 1.631E+00 7.483E-02
35 21 1.620E+00 1.103E-01 5.354E-02 1.631E+00 7.483E-02
36 20 1.618E+00 1.103E-01 5.341E-02 1.621E+00 7.394E-02
37 19 1.611E+00 1.099E-01 5.753E-02 1.621E+00 7.033E-02
38 18 1.574E+00 1.061E-01 6.485E-02 1.597E+00 6.006E-02
39 17 1.505E+00 1.020E-01 7.434E-02 1.526E+00 8.356E-02
40 16 1.411E+00 9.847E-02 8.116E-02 1.397E+00 8.976E-02
41 15 1.404E+00 9.814E-02 8.366E-02 1.397E+00 9.225E-02
42 11 1.404E+00 9.803E-02 8.323E-02 1.386E+00 9.006E-02
43 10 1.361E+00 9.533E-02 7.845E-02 1.355E+00 7.130E-02
44 8 1.221E+00 7.512E-02 3.500E-02 1.247E+00 4.389E-02
45++ 6 1.220E+00 7.148E-02 3.824E-02 1.203E+00 5.517E-02
46** 5 1.243E+00 7.171E-02 3.991E-02 1.263E+00 5.840E-02
47 4 1.281E+00 7.154E-02 4.064E-02 1.270E+00 4.950E-02
48 3 1.300E+00 7.015E-02 3.646E-02 1.319E+00 3.915E-02
49 2 1.325E+00 6.506E-02 2.839E-02 1.313E+00 3.556E-02
50 1 1.503E+00 5.698E-02 2.544E-02 1.484E+00 3.699E-02

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
* tree same as + tree
++ tree same as -- tree
+ tree same as ++ tree
* tree same as ++ tree
* tree same as -- tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Rel. risk is mean risk relative to sample average ignoring covariates

Cases fit give the number of cases used to fit node

Deviance is mean residual deviance for all cases in node

Wei-Yin Loh 184 GUIDE manual

5.9 Censored response: heart attack data 5 REGRESSION

Node Total Cases Matrix Node Node Split Other

label cases fit rank rel.risk deviance variable variables
1 500 500 1 1.000E+00 1.505E+00 age
2 244 244 1 3.726E-01 9.913E-01 chf
4T 49 49 1 1.110E+00 1.413E+00 miord
5T 195 195 1 2.124E-01 7.383E-01 year
3 256 256 1 1.890E+00 1.526E+00 chf
6T 106 106 1 3.028E+00 1.372E+00 sho
7 150 150 1 1.365E+00 1.469E+00 age
14T 120 120 1 1.063E+00 1.360E+00 los
15T 30 30 1 3.322E+00 1.278E+00 year

Number of terminal nodes of final tree: 5

Total number of nodes of final tree: 9
Second best split variable (based on curvature test) at root node is bmi

Regression tree:

Node 1: age <= 71.50000

Node 2: chf = "1"
Node 4: Risk relative to sample average ignoring covariates = 1.10956
Node 2: chf /= "1"
Node 5: Risk relative to sample average ignoring covariates = 0.21235
Node 1: age > 71.50000 or NA
Node 3: chf = "1"
Node 6: Risk relative to sample average ignoring covariates = 3.02760
Node 3: chf /= "1"
Node 7: age <= 85.50000 or NA
Node 14: Risk relative to sample average ignoring covariates = 1.06334
Node 7: age > 85.50000
Node 15: Risk relative to sample average ignoring covariates = 3.32215

***************************************************************

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse
variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if age <= 7.1500000E+01
age mean = 6.9846E+01
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val
Constant -3.5381E-02 -0.52 0.6041

Wei-Yin Loh 185 GUIDE manual

5.9 Censored response: heart attack data 5 REGRESSION

Predicted relative risk = 1.0000000000000000

----------------------------
Node 2: Intermediate node
A case goes into Node 4 if chf = "1"
chf mode = "0"
----------------------------
Node 4: Terminal node
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val
Constant 6.8580E-02 0.34 0.7332
Predicted relative risk = 1.1095574995429367
----------------------------
Node 5: Terminal node
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val
Constant -1.5849E+00 -7.43 0.0000
Predicted relative risk = 0.2123516881270062
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if chf = "1"
chf mode = "0"
----------------------------
Node 6: Terminal node
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val
Constant 1.0724E+00 9.89 0.0000
Predicted relative risk = 3.0276015801608627
----------------------------
Node 7: Intermediate node
A case goes into Node 14 if age <= 8.5500000E+01 or NA
age mean = 8.0667E+01
----------------------------
Node 14: Terminal node
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val
Constant 2.6029E-02 0.19 0.8459
Predicted relative risk = 1.0633351387096228
----------------------------
Node 15: Terminal node
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val
Constant 1.1652E+00 6.05 0.0000
Predicted relative risk = 3.3221527399879980
----------------------------

Observed and fitted values are stored in cons.fit

Wei-Yin Loh 186 GUIDE manual

5.9 Censored response: heart attack data 5 REGRESSION

age
71.50 1

chf chf
=1 2 =1 3

age
4 5 6 85.50 7
49 195 106
1.11 0.21 3.03
14 15
120 30
1.06 3.32
Figure 19: GUIDE v.26.0 0.50-SE piecewise constant relative risk regression tree for
predicting fstat. At each split, an observation goes to the left branch if and only
if the condition is satisfied. The symbol stands for or missing. Sample size
(in italics) and mean relative risks (relative to sample average ignoring covariates)
printed below nodes. Second best split variable at root node is bmi.

LaTeX code for tree is in cons.tex

The tree model, given in Figure 19, shows that risk of death is lowest (0.21
relative to the sample average for the whole data set) for those younger than 72 with
no congestive heart complications. The groups with the highest risks (3.033.32
relative to average) are those older than 71 with congestive heart complications and
those older than 85 without congestive heart complications.
The top few lines of the file whas500.fit and its column definitions are:
train node survivaltime logbasecumhaz relativerisk survivalprob mediansurvtime
y 14 2.178000E+03 -7.667985E-02 1.063335E+00 3.865048E-01 1.553833E+03
y 5 2.172000E+03 -7.667985E-02 2.123517E-01 8.270912E-01 2.354277E+03
y 5 2.190000E+03 -7.667985E-02 2.123517E-01 8.270912E-01 2.354277E+03
y 4 2.970000E+02 -1.320296E+00 1.109557E+00 7.512523E-01 1.534963E+03

The columns are:

train: y if the observation is used for model fitting, n if not.

Wei-Yin Loh 187 GUIDE manual

5.10 Multi-response: public health data 5 REGRESSION

node: terminal node label of observation.

survivaltime: observed survival time t.

logbasecumhaz:
Rt log of the estimated baseline cumulative hazard function log 0 (t) =
log 0 0 (u) du at observed time t.

relativerisk: exp( x ), risk of death relative to the average for the sample,
where x is the covariate vector of the observation, is the estimated regression
coefficient vector in the node, and is the coefficient in the constant model
0 (t) exp( ) fitted to all the training cases in the root node. Because a constant
is fitted to each node here, = 0.035381 is the value of at the root node.
For example, the first subject, which is in node 14, has = 0.026029 and so
relativerisk = exp( ) = exp(0.026029 + 0.035381) = 1.063335.

survivalprob: probability that the subject survives up to observed time t. For the
first subject, this is

exp{0 (t) exp( x)} = exp{ exp( + logbasecumhaz) relativerisk}

= exp( exp(0.035381 0.07667985) 1.063335)
= 0.3865049.

mediansurvtime: estimated median survival time t such that exp{0 (t) exp( x)} =
0.5, or, equivalently, 0 (t) exp( i x) = log(0.5), or logbasecumhaz(t) =
log log(2) i x, using linear interpolation of 0 (t). Median survival times
greater than the largest observed time have a trailing plus (+) sign.

5.10 Multi-response: public health data

GUIDE can fit a piecewise-constant regression model for two or more dependent
variables simultaneously (Loh and Zheng, 2013). We demonstrate this with the data
set phs.dat from a public health survey of about 120,000 respondents. There are
three D variables, namely, total restricted activity days in the past 2 week (raday),
number of doctor visits in the past 12 months (visit), and number of short-stay
hospital days in the past 12 months (hda12). The description files phs.dsc given
below lists 6 numeric and 9 categorical variables.
phs.dat
NA
1
1 phone c
2 sex c

Wei-Yin Loh 188 GUIDE manual

5.10 Multi-response: public health data 5 REGRESSION

3 age n
4 race c
5 marstat c
6 educ n
7 income n
8 poverty c
9 famsize n
10 condlist c
11 health n
12 latotal n
13 wkclass c
14 indus c
15 occup c
16 raday d
17 visit d
18 nacute x
19 hda12 d
20 lnvisit x

5.10.1 Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: mult.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: mult.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1):5
Option 5 is for multiresponse data.
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: phs.dsc
Reading data description file ...
Training sample file: phs.dat
Missing value code: NA

Wei-Yin Loh 189 GUIDE manual

5.10 Multi-response: public health data 5 REGRESSION

Records in data file start on line 1

Warning: N variables changed to S
Number of D variables = 3
D variables are:
raday
visit
hda12
Multivariate or univariate split variable selection:
Choose multivariate if there is an order among the D variables; otherwise choose univariate
Input 1 for multivariate, 2 for univariate ([1:2], <cr>=2):
We choose 2 because there is no order among the D variables.
Input 1 to normalize D variables, 2 for no normalization ([1:2], <cr>=1):
Input 1 for equal, 2 for unequal weighting of D variables ([1:2], <cr>=1):
Reading data file ...
Number of records in data file: 119579
Length of longest data entry: 17
Checking for missing values ...
Total number of cases: 119579
Missing values found among categorical variables
Separate categories will be created for missing categorical variables
Col. no. Categorical variable #levels #missing values
1 phone 4 0
2 sex 2 0
4 race 3 0
5 marstat 7 720
8 poverty 2 11400
10 condlist 6 754
13 wkclass 8 63793
14 indus 14 64059
15 occup 14 64085
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values
Finished processing 5000 of 119579 observations
Finished processing 10000 of 119579 observations
Finished processing 15000 of 119579 observations
Finished processing 20000 of 119579 observations
Finished processing 25000 of 119579 observations
Finished processing 30000 of 119579 observations
Finished processing 35000 of 119579 observations
Finished processing 40000 of 119579 observations
Finished processing 45000 of 119579 observations
Finished processing 50000 of 119579 observations
Finished processing 55000 of 119579 observations
Finished processing 60000 of 119579 observations
Finished processing 65000 of 119579 observations

Wei-Yin Loh 190 GUIDE manual

5.10 Multi-response: public health data 5 REGRESSION

Finished processing 70000 of 119579 observations

Finished processing 75000 of 119579 observations
Finished processing 80000 of 119579 observations
Finished processing 85000 of 119579 observations
Finished processing 90000 of 119579 observations
Finished processing 95000 of 119579 observations
Finished processing 100000 of 119579 observations
Finished processing 105000 of 119579 observations
Finished processing 110000 of 119579 observations
Finished processing 115000 of 119579 observations
Data checks complete
Normalizing data
Creating missing value indicators
Some D variables have missing values
You can use all the data or only those with complete D values
Using only cases with complete D values will reduce the sample size
but allows the option of using PCA for split selection
Input 1 to use all obs, 2 to use obs with complete D values ([1:2], <cr>=2):
Using only obs. with complete D values allows more options.
Rereading data
PCA can be used for variable selection
Do not use PCA if differential item functioning (DIF) scores are wanted
Input 1 to use PCA, 2 otherwise ([1:2], <cr>=2):
#cases w/ miss. D = number of cases with all D values missing
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
119579 0 30722 2 0 0 6 0 9
No. cases used for training: 60000
No. cases excluded due to 0 weight or missing D: 59579
Finished reading data file
Warning: interaction tests skipped
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): mult.tex
Input 2 to save node IDs of individual cases, 1 otherwise ([1:2], <cr>=2):
Input name of file to store terminal node ID of each case: mult.nid
Input 2 to save fitted values at each terminal node; 1 otherwise ([1:2], <cr>=1): 2
Input name of file to store node fitted values: mult.fit
Input file is created!
Run GUIDE with the command: guide < mult.in

5.10.2 Results
Multi-response or longitudinal data without T variables
Pruning by cross-validation
Data description file: phs.dsc

Wei-Yin Loh 191 GUIDE manual

5.10 Multi-response: public health data 5 REGRESSION

Training sample file: phs.dat

Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Number of D variables = 3
Univariate split variable selection method
Mean-squared errors (MSE) are calculated from normalized D variables
D variables equally weighted
Piecewise constant model
Number of records in data file: 119579
Length of longest data entry: 17
Missing values found among categorical variables
Separate categories will be created for missing categorical variables
Some D variables have missing values
Model fitted to subset of observations with complete D values
Neither LDA nor PCA used

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 phone c 4
2 sex c 2
3 age s 0.0000E+00 9.9000E+01
4 race c 3
5 marstat c 7 720
6 educ s 0.0000E+00 1.8000E+01 11162
7 income s 0.0000E+00 2.6000E+01 21547
8 poverty c 2 11400
9 famsize s 1.0000E+00 2.6000E+01
10 condlist c 6 754
11 health s 1.0000E+00 5.0000E+00 678
12 latotal s 1.0000E+00 4.0000E+00
13 wkclass c 8 63793
14 indus c 14 64059
15 occup c 14 64085
16 raday d 0.0000E+00 1.4000E+01
17 visit d 0.0000E+00 6.3700E+02 59579
19 hda12 d 0.0000E+00 2.8600E+02

#cases w/ miss. D = number of cases with all D values missing

Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
119579 0 30722 2 0 0 6 0 9
No. cases used for training: 60000

Wei-Yin Loh 192 GUIDE manual

5.10 Multi-response: public health data 5 REGRESSION

No. cases excluded due to 0 weight or missing D: 59579

Warning: interaction tests skipped

No interaction tests
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Split values for N and S variables based on exhaustive search
Max. number of split levels: 30
Min. node sample size: 3000
Number of SEs for pruned tree: 5.0000E-01

Size and CV Loss and SE of subtrees:

Tree #Tnodes Mean Loss SE(Mean) BSE(Mean) Median Loss BSE(Median)
1* 14 1.219E+00 7.364E-02 7.019E-02 1.180E+00 1.049E-01
2+ 12 1.219E+00 7.364E-02 7.021E-02 1.180E+00 1.050E-01
3 11 1.219E+00 7.364E-02 7.018E-02 1.180E+00 1.049E-01
4 9 1.219E+00 7.364E-02 7.017E-02 1.180E+00 1.049E-01
5 8 1.219E+00 7.364E-02 7.016E-02 1.180E+00 1.049E-01
6 7 1.219E+00 7.364E-02 7.016E-02 1.180E+00 1.049E-01
7 6 1.220E+00 7.364E-02 7.020E-02 1.181E+00 1.049E-01
8 5 1.221E+00 7.362E-02 7.017E-02 1.182E+00 1.048E-01
9 3 1.222E+00 7.363E-02 7.020E-02 1.183E+00 1.049E-01
10** 2 1.228E+00 7.370E-02 7.028E-02 1.189E+00 1.051E-01
11 1 1.310E+00 7.635E-02 7.183E-02 1.276E+00 1.051E-01

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree same as ++ tree
** tree same as -- tree
++ tree same as -- tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Cases fit give the number of cases used to fit node

MSE is residual sum of squares divided by number of cases in node
Node Total Cases Node Split
label cases fit MSE variable
1 60000 60000 6.571E-01 latotal
2T 5936 5936 3.844E+00 -
3T 54064 54064 2.615E-01 health

Wei-Yin Loh 193 GUIDE manual

5.10 Multi-response: public health data 5 REGRESSION

Number of terminal nodes of final tree: 2

Total number of nodes of final tree: 3
Second best split variable (based on curvature test) at root node is health

Regression tree for multi-response data:

Node 1: latotal <= 2.50000

Node 2: Mean cost = 3.84297E+00
Node 1: latotal > 2.50000 or NA
Node 3: Mean cost = 2.61520E-01

***************************************************************

In the following the predictor node mean is mean of complete cases.

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse
variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if latotal <= 2.5000000E+00
latotal mean = 3.7127E+00
Means of raday, visit, and hda12
6.1067E-01 4.1093E+00 6.0488E-01
The above 3 numbers are the mean values of the 3 D variables.
----------------------------
Node 2: Terminal node
Means of raday, visit, and hda12
2.8630E+00 1.2474E+01 3.0076E+00
----------------------------
Node 3: Terminal node
Means of raday, visit, and hda12
3.6337E-01 3.1909E+00 3.4108E-01
----------------------------

Case and node IDs are in file: mult.nid

LaTeX code for tree is in mult.tex

The tree is shown in Figure 20. The file mult.fit saves the mean values of the
dependent variables in each terminal node:
node raday visit hda12
2 0.28630E+01 0.12474E+02 0.30076E+01
3 0.36337E+00 0.31909E+01 0.34108E+00

Wei-Yin Loh 194 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

latotal
2.50 1

2 3
5936 54064
2.863 0.363
12.47 3.191
3.008 0.341
Figure 20: GUIDE v.26.0 0.50-SE regression tree for predicting response variables
raday, visit, and hda12. PCA not used. At each split, an observation goes to
the left branch if and only if the condition is satisfied. Sample sizes (in italics) and
predicted values of raday, visit, and hda12 printed below nodes. Second best split
variable at root node is health.

The file mult.nid gives the terminal node number for each observation, including
those that are not used to construct the tree (indicated by the letter n in the train
column of the file).

5.11 Longitudinal response with varying time: wage data

The data come from a longitudinal study on the hourly wage of 888 male high-
school dropouts (246 black, 204 Hispanic, 438 white), where the observation time
points as well as their number (113) varied across individuals (Murnane et al., 1999;
Singer and Willett, 2003). An earlier version of GUIDE was used to analyze the data
in Loh and Zheng (2013).
The response variable is hourly wage (in 1990 dollars) and the predictor variables
are hgc (highest grade completed; 612), exper (years in labor force; 0.00112.7
yrs), and race (Black, Hispanic, and White). The data file wagedat.txt is in wide
format, where each record refers to one individual. The description file wagedsc.txt
is given below. Note that observation time points are marked as t.
wagedat.txt
NA
1
1 id x
2 hgc n
3 exper1 t
4 exper2 t
5 exper3 t
6 exper4 t

Wei-Yin Loh 195 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

7 exper5 t
8 exper6 t
9 exper7 t
10 exper8 t
11 exper9 t
12 exper10 t
13 exper11 t
14 exper12 t
15 exper13 t
16 postexp1 x
17 postexp2 x
18 postexp3 x
19 postexp4 x
20 postexp5 x
21 postexp6 x
22 postexp7 x
23 postexp8 x
24 postexp9 x
25 postexp10 x
26 postexp11 x
27 postexp12 x
28 postexp13 x
29 wage1 d
30 wage2 d
31 wage3 d
32 wage4 d
33 wage5 d
34 wage6 d
35 wage7 d
36 wage8 d
37 wage9 d
38 wage10 d
39 wage11 d
40 wage12 d
41 wage13 d
42 ged1 x
43 ged2 x
44 ged3 x
45 ged4 x
46 ged5 x
47 ged6 x
48 ged7 x
49 ged8 x
50 ged9 x
51 ged10 x
52 ged11 x

Wei-Yin Loh 196 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

53 ged12 x
54 ged13 x
55 uerate1 x
56 uerate2 x
57 uerate3 x
58 uerate4 x
59 uerate5 x
60 uerate6 x
61 uerate7 x
62 uerate8 x
63 uerate9 x
64 uerate10 x
65 uerate11 x
66 uerate12 x
67 uerate13 x
68 race c

5.11.1 Input file creation

Because the default 0.5-SE rule yields a trivial tree with no splits, we show how the
options can be changed to produce a tree with the 0-SE rule. Following is a session
log.
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: wage.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: wage.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1): 2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1): 6
Input 1 for lowess smoothing, 2 for spline smoothing ([1:2], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1): 2
Choosing 1 will produce a 0.5-SE tree. We choose 2 to allow more options.
Input 1 for interaction tests, 2 to skip them ([1:2], <cr>=1):
Input 1 to prune by CV, 2 for no pruning ([1:2], <cr>=1):

Wei-Yin Loh 197 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: wagedsc.txt
Reading data description file ...
Training sample file: wagedat.txt
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Number of D variables = 13
Number of D variables = 13
D variables are:
wage1
wage2
wage3
wage4
wage5
wage6
wage7
wage8
wage9
wage10
wage11
wage12
wage13
T variables are:
exper1
exper2
exper3
exper4
exper5
exper6
exper7
exper8
exper9
exper10
exper11
exper12
exper13
D variables can be grouped into segments to look for patterns
Input 1 for roughly equal-sized groups, 2 for customized groups ([1:2], <cr>=1):
Input number of roughly equal-sized groups ([2:9], <cr>=3):
Input number of interpolating points for prediction ([10:100], <cr>=31):
Reading data file ...
Number of records in data file: 888
Length of longest data entry: 16
Checking for missing values ...
Total number of cases: 888

Wei-Yin Loh 198 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

Col. no. Categorical variable #levels #missing values

68 race 3 0
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values
Finished checking data
Creating missing value indicators
Rereading data
#cases w/ miss. D = number of cases with all D values missing
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
888 0 0 40 0 0 1 0 1
No. cases used for training: 888
No. cases excluded due to 0 weight or missing D: 0
Finished reading data file
Warning: interaction tests skipped
Default number of cross-validations = 10
Input 1 to accept the default, 2 to change it ([1:2], <cr>=1):
Best tree may be chosen based on mean or median CV estimate
Input 1 for mean-based, 2 for median-based ([1:2], <cr>=1):
Input number of SEs for pruning ([0.00:1000.00], <cr>=0.50): 0
This is where we choose the 0-SE pruning rule.
Choose a split point selection method for numerical variables:
Choose 1 to use faster method based on sample quantiles
Choose 2 to use exhaustive search
Input 1 or 2 ([1:2], <cr>=2):
Default max number of split levels = 10
Input 1 to accept this value, 2 to change it ([1:2], <cr>=1):
Default minimum node sample size is 44
Input 1 to use the default value, 2 to change it ([1:2], <cr>=1):
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): wage.tex
Input 1 to include node numbers, 2 to omit them ([1:2], <cr>=1):
Input 1 to number all nodes, 2 to number leaves only ([1:2], <cr>=1):
Choose a color for the terminal nodes:
(1) white
(2) lightgray
(3) gray
(4) darkgray
(5) black
(6) yellow
(7) red
(8) blue
(9) green
(10) magenta
(11) cyan

Wei-Yin Loh 199 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

Input your choice ([1:11], <cr>=6):

You can store the variables and/or values used to split and fit in a file
Choose 1 to skip this step, 2 to store split and fit variables,
3 to store split variables and their values
Input your choice ([1:3], <cr>=1): 3
Input file name: wage.var
Input 2 to save node IDs of individual cases, 1 otherwise ([1:2], <cr>=2):
Input name of file to store terminal node ID of each case: wage.nid
Input 2 to save fitted values at each terminal node; 1 otherwise ([1:2], <cr>=1): 2
Input name of file to store node fitted values: wage.fit
Input 2 to save terminal node IDs for importance scoring; 1 otherwise ([1:2], <cr>=1):
Input 2 to write R function for predicting new cases, 1 otherwise ([1:2], <cr>=1):
Input file is created!
Run GUIDE with the command: guide < wage.in

5.11.2 Results
Lowess smoothing
Longitudinal data with T variables
Pruning by cross-validation
Data description file: wagedsc.txt
Training sample file: wagedat.txt
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Number of D variables = 13
Number of D variables = 13
D variables are:
wage1
wage2
wage3
wage4
wage5
wage6
wage7
wage8
wage9
wage10
wage11
wage12
wage13
T variables are:
exper1
exper2
exper3

Wei-Yin Loh 200 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

exper4
exper5
exper6
exper7
exper8
exper9
exper10
exper11
exper12
exper13
Number of records in data file: 888
Length of longest data entry: 16
Model fitted to subset of observations with complete D values

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
2 hgc s 6.0000E+00 1.2000E+01
3 exper1 t 1.0000E-03 5.6370E+00
4 exper2 t 0.0000E+00 7.5840E+00 38
5 exper3 t 0.0000E+00 9.7770E+00 77
6 exper4 t 0.0000E+00 1.0815E+01 124
7 exper5 t 0.0000E+00 1.1777E+01 159
8 exper6 t 0.0000E+00 1.0587E+01 233
9 exper7 t 0.0000E+00 1.1279E+01 325
10 exper8 t 0.0000E+00 1.0582E+01 428
11 exper9 t 0.0000E+00 1.1621E+01 551
12 exper10 t 0.0000E+00 1.2260E+01 678
13 exper11 t 0.0000E+00 1.1980E+01 791
14 exper12 t 0.0000E+00 1.2558E+01 856
15 exper13 t 0.0000E+00 1.2700E+01 882
29 wage1 d 2.0299E+00 6.8649E+01
30 wage2 d 2.0689E+00 5.0400E+01 38
31 wage3 d 2.0462E+00 3.4501E+01 77
32 wage4 d 2.1170E+00 3.3149E+01 124
33 wage5 d 2.1043E+00 4.9304E+01 159
34 wage6 d 2.2078E+00 7.3995E+01 233
35 wage7 d 2.1043E+00 4.7276E+01 325
36 wage8 d 2.3164E+00 3.7713E+01 428
37 wage9 d 2.5294E+00 4.6109E+01 551
38 wage10 d 2.9982E+00 5.6543E+01 678
39 wage11 d 4.0837E+00 2.2198E+01 791
40 wage12 d 3.4315E+00 4.6201E+01 856
41 wage13 d 4.5631E+00 7.7757E+00 882

Wei-Yin Loh 201 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

68 race c 3

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
888 0 0 40 0 0 1 0 1
No. cases used for training: 888
No. cases excluded due to 0 weight or missing D: 0

Warning: interaction tests skipped

No interaction tests
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Split values for N and S variables based on exhaustive search
Max number of split levels = 10
Minimum node size = 44

Number of SEs for pruned tree = 0.0000E+00

Size and CV Loss and SE of subtrees:

Tree #Tnodes Mean Loss SE(Mean) BSE(Mean) Median Loss BSE(Median)
1 9 1.257E+02 1.045E+01 8.450E+00 1.204E+02 1.513E+01
2 7 1.257E+02 1.045E+01 8.450E+00 1.204E+02 1.513E+01
3 5 1.247E+02 1.053E+01 8.426E+00 1.185E+02 1.539E+01
4** 3 1.238E+02 1.053E+01 8.433E+00 1.175E+02 1.550E+01
5 2 1.239E+02 1.056E+01 8.631E+00 1.175E+02 1.562E+01
6++ 1 1.244E+02 1.064E+01 8.700E+00 1.157E+02 1.577E+01

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree same as -- tree
+ tree same as ++ tree
* tree same as ** tree
* tree same as -- tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Cases fit give the number of cases used to fit node

MSE is residual sum of squares divided by number of cases in node
Node Total Cases Node Split
label cases fit MSE variable
1 888 888 1.222E+02 hgc

Wei-Yin Loh 202 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

2T 577 577 1.040E+02 race

3 311 311 1.513E+02 race
6T 95 95 1.079E+02 -
7T 216 216 1.680E+02 hgc

Number of terminal nodes of final tree: 3

Total number of nodes of final tree: 5

Regression tree for longitudinal data:

Node 1: hgc <= 9.50000 or NA

Node 2: Mean cost = 1.03810E+02
Node 1: hgc > 9.50000
Node 3: race = "black"
Node 6: Mean cost = 1.06754E+02
Node 3: race /= "black"
Node 7: Mean cost = 1.67226E+02

***************************************************************

Node 1: Intermediate node

A case goes into Node 2 if hgc <= 9.5000000E+00 or NA
hgc mean = 8.9167E+00
----------------------------
Node 2: Terminal node
----------------------------
Node 3: Intermediate node
A case goes into Node 6 if race = "black"
race mode = "white"
----------------------------
Node 6: Terminal node
----------------------------
Node 7: Terminal node
----------------------------

Case and node IDs are in file: wage.nid

Node fitted values are in file: wage.fit
LaTeX code for tree is in wage.tex
Split and fit variable names are stored in wage.var

Figure 21 shows the tree and Figure 22 plots lowess-smoothed curves of mean
wage in the two terminal nodes. The plotting values are obtained from the result file
wage.fit whose contents are given below. The first column gives the node number
and the next two columns the start and end of the times at which fitted values are

Wei-Yin Loh 203 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

hgc
9.50 1

race
2 = black 3
577

6 7
95 216

Figure 21: GUIDE v.26.0 0.00-SE regression tree for predicting longitudinal variables
wage1, wage2, etc. At each split, an observation goes to the left branch if and only
if the condition is satisfied. The symbol stands for or missing. For splits
on categorical variables, values not present in the training sample go to the right.
Sample sizes (in italics) printed below nodes.

computed. The other columns give the fitted values equally spaced between the start
and end times.

node t.start t.end fitted1 fitted2 fitted3 fitted4 fitted5 fitted6 fitted7 fitted8 fitted9 fitted10
2 0.10000E-02 0.12700E+02 0.48875E+01 0.51221E+01 0.53241E+01 0.54668E+01 0.55738E+01 0
6 0.80000E-02 0.12558E+02 0.61270E+01 0.58648E+01 0.57522E+01 0.57674E+01 0.57653E+01 0
7 0.20000E-02 0.12045E+02 0.56786E+01 0.58892E+01 0.60859E+01 0.62420E+01 0.63533E+01 0

The file wage.var below gives the type (t if node is terminal) and name of the
variable used to split each node and the split point (for n or s variables) or split
values (if c variable). The word NONE indicates a terminal node that cannot be split
by any variable. For a non-terminal node, the integer in the 5th column indicates
the number of split values to follow on the line.
1 s hgc hgc 1 0.9500000000E+01
2 t race race 0.0000000000E+00
3 c race race 1 "black"
6 t NONE NONE 0.0000000000E+00
3 c race race 1 "black"
7 t hgc hgc 0.0000000000E+00

Wei-Yin Loh 204 GUIDE manual

5.11 Longitudinal response with varying time: wage data 5 REGRESSION

hgc 9
11

black, hgc > 9

nonblack, hgc > 9
10
hourly wage ($)

9
8
7
6
5

0 2 4 6 8 10 12

exper (years)

Figure 22: Lowess-smoothed mean wage curves in the terminal nodes of Figure 21.

Wei-Yin Loh 205 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

5.12 Subgroup identification: breast cancer

GUIDE has several methods to identify subgroups for differential treatment effects
from randomized experiments. See Loh et al. (2015) and Loh et al. (2016) for details.
The treatment variable is assumed to be categorical (i.e., it takes nominal values)
and the response is an uncensored or censored event time (e.g., survival time). The
key points are:

1. The treatment variable is designated as R (for Rx).

2. If there is no censoring in the response, it is designated as the dependent

variable as D as usual.

3. If there is censoring in the response, the variable is designated as T (first letter

of Time). In this case, the event indicator is designated as D (first letter of
Death) and takes value 1 if the event (death) occurs and 0 if the event
time is censored.

There are two types of covariate variables in subgroup identification. A prognos-

tic variable is a clinical or biologic characteristic that provides information on the
likely outcome of the disease in an untreated individual (e.g., patient age, family
history, disease stage, and prior therapy). A predictive variable is a characteristic
that provides information on the likely benefit from treatment. Predictive variables
can be used to identify subgroups of patients who are most likely to benefit from a
given therapy. Therefore prognostic variables define the effects of patient or tumor
characteristics on the patient outcome, whereas predictive variables define the effect
of treatment on the tumor (Italiano, 2011). Accordingly, GUIDE has two methods,
called Gi and Gs. Gi is more sensitive to predictive variables and Gs tends to be
equally sensitive to prognostic and predictive variables (Loh et al., 2015).

5.12.1 Without linear prognostic control

The simplest model only uses the covariates to split the intermediate nodes; termi-
nal nodes are fitted with treatment means. We use a data set from a randomized
controlled breast cancer trial (Schmoor et al., 1996) to show this. The data are in
the file cancer.txt; it can also be obtained from the TH.data R package (Hothorn,
2017). In the description file cancerdsc.txt below, the treatment variable is hor-
mone therapy, horTh. The variable time is (censored) time to recurrence of cancer
and event = 1 if the cancer recurred and = 0 if it did not. Ordinal predictor variables
may be designated as n or s (with this option of no linear prognostic control, n
variables will be automatically changed to s when the program is executed).

Wei-Yin Loh 206 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

cancer.txt
NA
1
1 horTh r
2 age n
3 menostat c
4 tsize n
5 tgrade c
6 pnodes n
7 progrec n
8 estrec n
9 time t
10 event d

Input file generation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: nolin.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: nolin.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1): 2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1): 4
Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 2 for linear prognostic control,
choose 3 for no linear prognostic control
1: multiple linear, 2: best simple linear, 3: constant ([1:3], <cr>=3):
Option 3 is the one for no linear prognostic control.
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: cancerdsc.txt
Reading data description file ...
Training sample file: cancer.txt
Missing value code: NA

Wei-Yin Loh 207 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

Records in data file start on line 1

R variable present
Warning: N variables changed to S
Warning: model changed to linear in treatment
Warnings due to presence of R variable and choice of no linear prognostic effects.
Dependent variable is death
Reading data file ...
Number of records in data file: 686
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 686
Col. no. Categorical variable #levels #missing values
1 horTh 2 0
3 menostat 2 0
5 tgrade 3 0
Re-checking data ...
Assigning codes to categorical and missing values
Finished checking data
Smallest uncensored T: 72.00
No. cases dropped due to missing D or T or censored T < smallest uncensored T: 14
No. complete cases excluding censored T < smallest uncensored T: 672
No. cases used to compute baseline hazard: 672
No. cases with D=1 and T >= smallest uncensored: 299
GUIDE will try to create the variables in the description file.
If it is unsuccessful, please create the columns yourself...
Number of dummy variables created: 1
Choose a subgroup identification method:
1 = Sum of chi-squares (Gs)
2 = Treatment interactions (Gi)
Input your choice: ([1:2], <cr>=2):
Gi is the choice if splitting on predictive variables is preferred.
Creating dummy variables
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var #R-var
686 0 0 0 0 0 5 0 2 1
Survival time variable in column: 9
Event indicator variable in column: 10
Proportion uncensored among nonmissing T and D variables = 0.445
No. cases used for training: 672
Finished reading data file
Warning: interaction tests skipped
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): nolin.tex
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: nolin.fit

Wei-Yin Loh 208 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

Input file is created!

Run GUIDE with the command: guide < nolin.in

Results The contents of nolin.out follow.

Proportional hazards regression with relative risk estimates
Pruning by cross-validation
Data description file: cancerdsc.txt
Training sample file: cancer.dat
Missing value code: NA
Records in data file start on line 1
R variable present
Dependent variable is event
Piecewise linear model
Number of records in data file: 686
Length of longest data entry: 4
Smallest uncensored T: 72.00
No. cases dropped due to missing D or T or censored T < smallest uncensored T: 14
No. complete cases excluding censored T < smallest uncensored T: 672
No. cases used to compute baseline hazard: 672
No. cases with D=1 and T >= smallest uncensored: 299
Number of dummy variables created: 1

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
t=survival time variable
Column Name Minimum Maximum #Categories #Missing
1 horTh r 2
2 age s 2.1000E+01 8.0000E+01
3 menostat c 2
4 tsize s 3.0000E+00 1.2000E+02
5 tgrade c 3
6 pnodes s 1.0000E+00 5.1000E+01
7 progrec s 0.0000E+00 2.3800E+03
8 estrec s 0.0000E+00 1.1440E+03
9 time t 7.2000E+01 2.6590E+03
10 event d 0.0000E+00 1.0000E+00
=================== Constructed variables ===================
11 lnbasehaz z -6.5103E+00 5.8866E-02
12 horTh.yes f 0.0000E+00 1.0000E+00

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var

Wei-Yin Loh 209 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

686 0 0 0 0 0 5 0 2
Survival time variable in column: 9
Event indicator variable in column: 10
Proportion uncensored among nonmissing T and D variables: 0.445
No. cases used for training: 672

Warning: interaction tests skipped

Missing values imputed with node means for regression
Treatment interactions (Gi)
No interaction tests
Pruning by v-fold cross-validation, with v = 10
Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node: 1.0000
Max. number of split levels: 10
Min. number of cases per treatment at each node: 2
Min. node sample size: 33
Number of iterations: 5
Number of SEs for pruned tree: 5.0000E-01

Size and CV Loss and SE of subtrees:

Tree #Tnodes Mean Loss SE(Mean) BSE(Mean) Median Loss BSE(Median)
1 14 1.441E+00 5.809E-02 5.898E-02 1.392E+00 5.401E-02
2 13 1.439E+00 5.802E-02 5.934E-02 1.392E+00 5.413E-02
3 12 1.439E+00 5.802E-02 5.934E-02 1.392E+00 5.413E-02
4 11 1.431E+00 5.780E-02 6.052E-02 1.392E+00 6.089E-02
5 9 1.427E+00 5.757E-02 6.059E-02 1.388E+00 5.826E-02
6 8 1.427E+00 5.757E-02 6.059E-02 1.388E+00 5.826E-02
7+ 7 1.422E+00 5.706E-02 5.355E-02 1.388E+00 5.422E-02
8++ 4 1.413E+00 5.498E-02 4.407E-02 1.394E+00 3.635E-02
9** 2 1.426E+00 5.214E-02 3.733E-02 1.434E+00 4.886E-02
10 1 1.442E+00 5.157E-02 1.216E-02 1.450E+00 1.474E-02

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree same as -- tree
* tree same as ++ tree

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Rel. risk is mean risk relative to sample average ignoring covariates

Wei-Yin Loh 210 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

Cases fit give the number of cases used to fit node

Deviance is mean residual deviance for all cases in node
Node Total Cases Matrix Node Node Split
label cases fit rank rel.risk deviance variable
1 672 672 1 1.000E+00 1.414E+00 progrec
2T 274 274 1 1.588E+00 1.584E+00 estrec
3T 398 398 1 7.095E-01 1.172E+00 menostat

Number of terminal nodes of final tree: 2

Total number of nodes of final tree: 3
Second best split variable (based on curvature test) at root node is estrec

Regression tree:

Node 1: progrec <= 21.50000

Node 2: Risk relative to sample average ignoring covariates = 1.58824
Node 1: progrec > 21.50000 or NA
Node 3: Risk relative to sample average ignoring covariates = 0.70947

***************************************************************

Constant term for constant hazard model (ignoring covariates): -0.00259998

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse
variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if progrec <= 2.1500000E+01
progrec mean = 1.1092E+02
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 1.2903E-01 1.85 0.0651
horTh.yes -3.6984E-01 -2.97 0.0031 0.0000E+00 3.6012E-01 1.0000E+00
----------------------------
Node 2: Terminal node
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 5.0439E-01 5.04 0.0000
horTh.yes -1.1775E-01 -0.71 0.4786 0.0000E+00 3.6131E-01 1.0000E+00
----------------------------
Node 3: Terminal node
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum

Wei-Yin Loh 211 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

progrec
21.50 1

1.660 0.879
1.476 2 3 0.459
274 398

Figure 23: GUIDE v.26.0 0.50-SE Gi proportional hazards regression tree for differen-
tial treatment effects without linear prognostic control. At each split, an observation
goes to the left branch if and only if the condition is satisfied. Numbers beside
terminal nodes are estimated relative risks (relative to average for sample ignoring
covariates) corresponding to horTh levels no and yes, respectively; numbers (in ital-
ics) below are sample sizes. Nodes with negative and positive effects for horTh level
yes are colored red and green, respectively. Second best split variable at root node
is estrec.

Constant -1.3184E-01 -1.35 0.1775

horTh.yes -6.5011E-01 -3.40 0.0007 0.0000E+00 3.5930E-01 1.0000E+00
----------------------------

Observed and fitted values are stored in nolin.fit

LaTeX code for tree is in nolin.tex

Let (u, x) denote the hazard function at time u and predictor values x and
let 0 (u) denote the baseline hazard function. The results show that the fitted
proportional hazards model is

(u, x) = 0 (u)[exp{1 + 1 I(horTh = yes)}I(progrec 21.5)

+ exp{2 + 2 I(horTh = yes)}I(progrec > 21.5)]

with 1 = 0.50439, 1 = 0.11775, 2 = 0.13184, and 2 = 0.65011.

Figure 23 shows the LATEX tree diagram. The numbers beside each terminal
node are relative risks (relative to the average risk of the entire sample) defined
as exp{ + I(horTh = yes) }, where = 0.00259998 (printed in the line
below the row of asterisks in the results) is the estimated regression coefficient for
the constant model (u, x) = 0 (u) exp( ) fitted to the entire sample (see the text
in Section 5.9 on page 187). For example, the relative risks for horTh = no and yes

Wei-Yin Loh 212 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

Node 2 Node 3
Survival probability

0.8

0.8
0.4

0.4
horTh = yes horTh = yes
horTh = no horTh = no
0.0

0.0
500 1000 2000 500 1000 2000

Figure 24: Estimated survival probability functions for breast cancer data

in the left terminal node of the tree are

exp(0.50439 + 0.00259998) = 1.660286
exp(0.50439 0.11775 + 0.00259998) = 1.475859
respectively. The Kaplan-Meier survival functions estimated from the data in the
terminal nodes of the tree are shown in Figure 24.

Estimated relative risks and survival probabilities The file nolin.fit gives
the terminal node number, estimated survival time, log baseline cumulative haz-
ard, relative risk (relative to the average for the data, ignoring covariates), survival
probability, and median survival time of each observation in the training sample
file cancer.txt. The results for the first few observations are shown below. See
Section 5.9 for definitions of the terms.
train node survivaltime logbasecumhaz relativerisk survivalprob mediansurvtime
y 3 1.814000E+03 -3.317667E-01 8.787636E-01 5.331186E-01 2.014420E+03
y 3 2.018000E+03 -2.024282E-01 4.587030E-01 6.882035E-01 2.659000E+03+
y 3 7.120000E+02 -1.300331E+00 4.587030E-01 8.828100E-01 2.659000E+03+
y 3 1.807000E+03 -3.550694E-01 4.587030E-01 7.255880E-01 2.659000E+03+
y 3 7.720000E+02 -1.176558E+00 8.787636E-01 7.631865E-01 2.014420E+03
y 2 4.480000E+02 -2.105688E+00 1.660293E+00 8.173929E-01 1.038277E+03

5.12.2 With linear prognostic control

To reduce or eliminate confounding between treatment and covariate variables, it
may be desirable to adjust for the effects of the latter by fitting a regression model

Wei-Yin Loh 213 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

that includes the most important covariate as a linear predictor in each node. This
is accomplished by choosing best simple polynomial option and specifying each
potential linear predictor as n in the description file (no change is needed in
cancerdsc.txt).

Input file generation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: lin.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: lin.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1): 2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1): 4
Choose complexity of model to use at each node:
If R variable present (i.e., subgroup identification),
choose 1 for multiple regression (including R var.) in each node,
choose 2 to fit one linear prognostic var. (N or F) in each node,
choose 3 (constant) to fit only treatment effect in each node
1: multiple linear, 2: best simple linear, 3: constant ([1:3], <cr>=3): 2
Option 2 activates linear prognostic control.
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: cancerdsc.txt
Reading data description file ...
Training sample file: cancer.txt
Missing value code: NA
Records in data file start on line 1
R variable present
Dependent variable is death
Reading data file ...
Number of records in data file: 686
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 686

Wei-Yin Loh 214 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

Col. no. Categorical variable #levels #missing values

1 horTh 2 0
3 menostat 2 0
5 tgrade 3 0
Re-checking data ...
Assigning codes to categorical and missing values
Data checks complete
Smallest uncensored T: 72.00
No. cases dropped due to missing D or T or censored T < smallest uncensored T: 14
No. complete cases excluding censored T < smallest uncensored T: 672
No. cases used to compute baseline hazard: 672
No. cases with D=1 and T >= smallest uncensored: 299
GUIDE will try to create the variables in the description file.
If it is unsuccessful, please create the columns yourself...
Number of dummy variables created: 1
Choose a subgroup identification method:
1 = Sum of chi-squares (Gs)
2 = Treatment interactions (Gi)
Input your choice: ([1:2], <cr>=2):
Creating dummy variables
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var #R-var
686 0 0 0 5 0 0 0 2 1
Survival time variable in column: 9
Event indicator variable in column: 10
Proportion uncensored among nonmissing T and D variables: .445
No. cases used for training: 672
Finished reading data file
Warning: interaction tests skipped
Input 1 for LaTeX tree code, 2 to skip it ([1:2], <cr>=1):
Input file name to store LaTeX code (use .tex as suffix): lin.tex
Input 2 to save individual fitted values and node IDs, 1 otherwise ([1:2], <cr>=2):
Input name of file to store node ID and fitted value of each case: lin.fit
Input file is created!
Run GUIDE with the command: guide < lin.in

Results The contents of the output file lin.out follows.

Proportional hazards regression with relative risk estimates
No truncation of predicted values
Pruning by cross-validation
Data description file: cancerdsc.txt
Training sample file: cancer.dat
Missing value code: NA

Wei-Yin Loh 215 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

Records in data file start on line 1

R variable present
Dependent variable is event
Piecewise simple linear or constant model
Powers are dropped if they are not significant at level 1.0000
Number of records in data file: 686
Length of longest data entry: 4
Smallest uncensored T: 72.00
No. cases dropped due to missing D or T or censored T < smallest uncensored T: 14
No. complete cases excluding censored T < smallest uncensored T: 672
No. cases used to compute baseline hazard: 672
No. cases with D=1 and T >= smallest uncensored: 299
Number of dummy variables created: 1

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
t=survival time variable
Column Name Minimum Maximum #Categories #Missing
1 horTh r 2
2 age n 2.1000E+01 8.0000E+01
3 menostat c 2
4 tsize n 3.0000E+00 1.2000E+02
5 tgrade c 3
6 pnodes n 1.0000E+00 5.1000E+01
7 progrec n 0.0000E+00 2.3800E+03
8 estrec n 0.0000E+00 1.1440E+03
9 time t 7.2000E+01 2.6590E+03
10 event d 0.0000E+00 1.0000E+00
=================== Constructed variables ===================
11 lnbasehaz z -6.5103E+00 5.8866E-02
12 horTh.yes f 0.0000E+00 1.0000E+00

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
686 0 0 0 5 0 0 0 2
Survival time variable in column: 9
Event indicator variable in column: 10
Proportion uncensored among nonmissing T and D variables: 0.445
No. cases used for training: 672

Warning: interaction tests skipped

Missing values imputed with node means for regression
Treatment interactions (Gi)
No interaction tests

Wei-Yin Loh 216 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

Pruning by v-fold cross-validation, with v = 10

Selected tree is based on mean of CV estimates
Fraction of cases used for splitting each node: 1.0000
Max. number of split levels: 10
Min. number of cases per treatment at each node: 2
Min. node sample size: 33
Number of iterations: 5
Number of SEs for pruned tree: 5.0000E-01

Size and CV Loss and SE of subtrees:

Tree #Tnodes Mean Loss SE(Mean) BSE(Mean) Median Loss BSE(Median)
1 13 1.556E+00 7.034E-02 6.171E-02 1.514E+00 5.084E-02
2 12 1.537E+00 6.844E-02 6.134E-02 1.514E+00 4.839E-02
3 11 1.532E+00 6.763E-02 5.804E-02 1.503E+00 5.154E-02
4 10 1.522E+00 6.671E-02 5.404E-02 1.503E+00 5.478E-02
5 9 1.517E+00 6.622E-02 4.895E-02 1.513E+00 5.644E-02
6 7 1.502E+00 6.478E-02 4.874E-02 1.513E+00 6.413E-02
7 5 1.445E+00 6.189E-02 6.038E-02 1.391E+00 8.305E-02
8 4 1.426E+00 6.029E-02 5.445E-02 1.347E+00 5.465E-02
9** 2 1.368E+00 5.303E-02 4.107E-02 1.333E+00 3.131E-02
10 1 1.427E+00 5.729E-02 2.922E-02 1.406E+00 3.550E-02

0-SE tree based on mean is marked with *

0-SE tree based on median is marked with +
Selected-SE tree based on mean using naive SE is marked with **
Selected-SE tree based on mean using bootstrap SE is marked with --
Selected-SE tree based on median and bootstrap SE is marked with ++
** tree and ++ tree are the same

Following tree is based on mean CV with naive SE estimate (**).

Structure of final tree. Each terminal node is marked with a T.

Rel. risk is mean risk relative to sample average ignoring covariates

Cases fit give the number of cases used to fit node

Deviance is mean residual deviance for all cases in node
Node Total Cases Matrix Node Node Split
label cases fit rank rel.risk deviance variable
1 672 672 3 1.000E+00 1.392E+00 progrec
2T 292 292 3 1.546E+00 1.519E+00 estrec
3T 380 380 3 6.998E-01 1.155E+00 menostat

Number of terminal nodes of final tree: 2

Total number of nodes of final tree: 3
Second best split variable (based on curvature test) at root node is estrec

Wei-Yin Loh 217 GUIDE manual

5.12 Subgroup identification: breast cancer 5 REGRESSION

Regression tree:

Node 1: progrec <= 24.50000

Node 2: Risk relative to sample average ignoring covariates = 1.54577
Node 1: progrec > 24.50000 or NA
Node 3: Risk relative to sample average ignoring covariates = 0.69983

***************************************************************

Constant term for constant hazard model (ignoring covariates): -0.01375697

WARNING: p-values below not adjusted for split search. For a bootstrap
solution, see Loh et al. (2016), "Identification of subgroups with
differential treatment effects for longitudinal and multiresponse
variables", Statistics in Medicine, v.35, 4837-4855.

Node 1: Intermediate node

A case goes into Node 2 if progrec <= 2.4500000E+01
progrec mean = 1.1092E+02
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant -1.7811E-01 -2.15 0.0322
pnodes 5.8599E-02 8.99 0.0000 1.0000E+00 4.9866E+00 5.1000E+01
horTh.yes -3.6299E-01 -2.91 0.0037 0.0000E+00 3.6012E-01 1.0000E+00
----------------------------
Node 2: Terminal node
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant 9.9061E-03 0.08 0.9336
pnodes 8.6795E-02 8.35 0.0000 1.0000E+00 5.7089E+00 3.6000E+01
horTh.yes -2.0916E-01 -1.27 0.2063 0.0000E+00 3.4932E-01 1.0000E+00
----------------------------
Node 3: Terminal node
Coefficients of log-relative risk function:
Regressor Coefficient t-stat p-val Minimum Mean Maximum
Constant -3.3455E-01 -2.81 0.0051
pnodes 3.9899E-02 3.61 0.0003 1.0000E+00 4.4316E+00 5.1000E+01
horTh.yes -6.4331E-01 -3.30 0.0011 0.0000E+00 3.6842E-01 1.0000E+00
----------------------------

Observed and fitted values are stored in lin.fit

LaTeX code for tree is in lin.tex

Wei-Yin Loh 218 GUIDE manual

 6 IMPORTANCE SCORING

progrec
24.50 1

2 3
292 380
1.55 0.70
pnodes pnodes

Figure 25: GUIDE v.26.0 0.50-SE Gi proportional hazards regression tree for differ-
ential treatment effects with linear prognostic control. At each split, an observation
goes to the left branch if and only if the condition is satisfied. Sample size (in italics),
relative risks (relative to average for sample ignoring covariates), and name of linear
prognostic variable printed below nodes. Second best split variable at root node is
estrec.

Without controlling for split selection, the p-value of the effect of treatment
(horTh) is 0.0011 after adjustment with prognostic variable pnodes in terminal
node 3. The small size of the p-value suggests that the treatment effect in the node
will remain significant at the 0.05 level after controlling for split selection (one ap-
proach to adjusting for split selection is the bootstrap method proposed in Loh et al.
(2016)). The treatment effect in node 2, however, is not significant even without
controlling for split selection. The tree is shown in Figure 25.

6 Importance scoring
When there are numerous predictor variables, it may be useful to rank them in order
of their importance. GUIDE has a facility to do this. In addition, it provides a
threshold for distinguishing the important variables from the unimportant onessee
Loh et al. (2015) and Loh (2012); the latter also shows that using GUIDE to find a
subset of variables can increase the prediction accuracy of a model.

6.1 Classification: glaucoma data

6.1.1 Input file creation
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file

Wei-Yin Loh 219 GUIDE manual

6.1 Classification: glaucoma data 6 IMPORTANCE SCORING

Input your choice: 1

Name of batch input file: imp.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):2
Name of batch output file: imp.out
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: glaucoma.dsc
Reading data description file ...
Training sample file: glaucoma.rdata
Missing value code: NA
Records in data file start on line 2
Warning: N variables changed to S
Dependent variable is Class
Reading data file ...
Number of records in data file: 170
Length of longest data entry: 8
Checking for missing values ...
Total number of cases: 170
Number of classes = 2
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values
Finished checking data
Creating missing value indicators
Rereading data
Class #Cases Proportion
glaucoma 85 0.50000000
normal 85 0.50000000
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
170 0 17 0 0 0 66 0 0
No. cases used for training: 170
No. cases excluded due to 0 weight or missing D: 0
Finished reading data file
Input expected fraction of noise variables erroneously selected ([0.00:0.99], <cr>=0.01):
Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3 ([1:3], <cr>=2):
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1):
You can create a description file with the selected variables included or excluded
Input 2 to create such a file, 1 otherwise ([1:2], <cr>=1):

Wei-Yin Loh 220 GUIDE manual

6.1 Classification: glaucoma data 6 IMPORTANCE SCORING

You can also output the importance scores and variable names to a file
Input 1 to create such a file, 2 otherwise ([1:2], <cr>=1):
Input file name: imp.scr
Input file is created!
Run GUIDE with the command: guide < imp.in

6.1.2 Contents of imp.out

The most interesting part of the output file is at the end which, for this data set,
is given below. The variables are sorted according to their importance scores, with
a cut-off value of 1.0 separating the potentially important variables from the unim-
portant onessee Loh (2012) and Loh et al. (2015) for details.
Predictor variables sorted by importance scores
Importance Scores
Scaled Unscaled Rank Variable
100.0 1.06846E+01 1.00 clv
86.6 9.25005E+00 2.00 lora
71.9 7.68730E+00 3.00 vars
68.7 7.33771E+00 4.00 vari
67.4 7.20533E+00 5.00 varg
61.6 6.58214E+00 6.00 rnf
58.3 6.22521E+00 7.00 tmg
56.7 6.06296E+00 8.00 tmi
54.0 5.76561E+00 9.00 varn
53.5 5.72145E+00 10.00 phcn
52.4 5.59894E+00 11.00 vbri
51.7 5.52275E+00 12.00 cs
49.4 5.27423E+00 13.00 vart
49.1 5.24279E+00 14.00 abri
47.1 5.03325E+00 15.00 hic
46.7 4.99261E+00 16.00 tms
44.7 4.78042E+00 17.00 abrs
43.8 4.68273E+00 18.00 phcg
42.9 4.58742E+00 19.00 abrg
41.4 4.42721E+00 20.00 mhcn
38.6 4.12213E+00 21.00 phci
36.7 3.92249E+00 22.00 abrn
36.1 3.86040E+00 23.00 vbrg
35.4 3.78652E+00 24.00 vbrn
33.8 3.60679E+00 25.00 ean
33.5 3.57969E+00 26.00 vbrs
33.3 3.55618E+00 27.00 mhci
32.2 3.44047E+00 28.00 mdic
30.0 3.20232E+00 29.00 mhcg

Wei-Yin Loh 221 GUIDE manual

6.1 Classification: glaucoma data 6 IMPORTANCE SCORING

29.6 3.15883E+00 30.00 vbrt

29.1 3.11387E+00 31.00 eai
28.5 3.04736E+00 32.00 vbsn
28.1 2.99958E+00 33.00 hvc
27.9 2.97739E+00 34.00 vbsi
27.3 2.91206E+00 35.00 tmt
26.0 2.77923E+00 36.00 mhcs
25.5 2.72678E+00 37.00 abrt
24.0 2.56639E+00 38.00 vbsg
23.2 2.47431E+00 39.00 vbss
21.5 2.29890E+00 40.00 phcs
21.0 2.24008E+00 41.00 vasi
19.9 2.12300E+00 42.00 emd
19.8 2.11122E+00 43.00 vass
19.5 2.08129E+00 44.00 eag
18.9 2.02262E+00 45.00 vasg
14.9 1.59099E+00 46.00 eas
13.0 1.38862E+00 47.00 vbst
12.5 1.33304E+00 48.00 vasn
12.2 1.30346E+00 49.00 tmn
11.9 1.27316E+00 50.00 eat
10.8 1.15190E+00 51.00 vast
9.9 1.05271E+00 52.00 at
----------------- cut-off -------------------
8.6 9.22470E-01 53.00 mdn
7.2 7.67824E-01 54.00 ai
5.7 6.13148E-01 55.00 mdg
5.4 5.79434E-01 56.00 an
5.0 5.29275E-01 57.00 mds
4.5 4.77615E-01 58.00 mhct
4.0 4.27601E-01 59.00 mdi
4.0 4.23403E-01 60.00 ag
3.9 4.16006E-01 61.00 mr
2.7 2.91827E-01 62.00 as
2.6 2.80042E-01 63.00 mdt
2.5 2.70427E-01 64.00 tension
2.4 2.56020E-01 65.00 mv
2.1 2.22411E-01 66.00 phct
Variables with unscaled scores above 1 are important

Number of important and unimportant split variables: 52, 14

The scores are also printed in the file imp.scr. Following is an R file for graphing
them in Figure 26.
z0 <- read.table("imp.scr",header=TRUE)

Wei-Yin Loh 222 GUIDE manual

6.2 Regression with censoring: heart attack data 6 IMPORTANCE SCORING

par(mar=c(5,6,2,1),las=1)
barplot(z0$Score,names.arg=z0$Variable,col="cyan",horiz=TRUE,xlab="Importance scores")
abline(v=1,col="red",lty=2)

6.2 Regression with censoring: heart attack data

We show how to obtain the importance scores for the Worcester Heart Attack Study
data analyzed in Section 5.9.
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: whas500imp.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1): 2
Name of batch output file: whas500imp.out
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1): 2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1): 4
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: whas500.dsc
Reading data description file ...
Training sample file: whas500.csv
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is fstat
Reading data file ...
Number of records in data file: 500
Length of longest data entry: 10
Checking for missing values ...
Total number of cases: 500
Col. no. Categorical variable #levels #missing values
3 gender 2 0
8 cvd 2 0
9 afb 2 0
10 sho 2 0

Wei-Yin Loh 223 GUIDE manual

6.2 Regression with censoring: heart attack data 6 IMPORTANCE SCORING

phct
mv
tension
mdt
as
mr
ag
mdi
mhct
mds
an
mdg
ai
mdn
at
vast
eat
tmn
vasn
vbst
eas
vasg
eag
vass
emd
vasi
phcs
vbss
vbsg
abrt
mhcs
tmt
vbsi
hvc
vbsn
eai
vbrt
mhcg
mdic
mhci
vbrs
ean
vbrn
vbrg
abrn
phci
mhcn
abrg
phcg
abrs
tms
hic
abri
vart
cs
vbri
phcn
varn
tmi
tmg
rnf
varg
vari
vars
lora
clv

0 2 4 6 8 10

Importance scores

Figure 26: Importance scores for glaucoma data; variables with bars shorter than
indicated by the red dashed line are considered unimportant.

Wei-Yin Loh 224 GUIDE manual

6.2 Regression with censoring: heart attack data 6 IMPORTANCE SCORING

11 chf 2 0
12 av3 2 0
13 miord 2 0
14 mitype 2 0
15 year 3 0
Re-checking data ...
Assigning codes to categorical and missing values
Data checks complete
Smallest uncensored T: 1.0000
No. complete cases excluding censored T < smallest uncensored T: 500
No. cases used to compute baseline hazard: 500
No. cases with D=1 and T >= smallest uncensored: 215
Rereading data
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
500 0 0 5 0 0 6 0 9
Survival time variable in column: 21
Event indicator variable in column: 22
Proportion uncensored among nonmissing T and D variables: .430
No. cases used for training: 500
Finished reading data file
Input expected fraction of noise variables erroneously selected ([0.00:0.99], <cr>=0.01):
You can create a description file with the selected variables included or excluded
Input 2 to create such a file, 1 otherwise ([1:2], <cr>=1): 2
Input 1 to keep only selected variables, 2 to exclude selected variables ([1:2], <cr>=1):
Input file name: whas500new.dsc
You can also output the importance scores and variable names to a file
Input 1 to create such a file, 2 otherwise ([1:2], <cr>=1):
Input file name: whas500imp.scr
Input file is created!
Run GUIDE with the command: guide < whas500imp.in

Results The importance scores are given at the end of the output file whas500imp.out.
Variables with scores below 1.0 (i.e., below the cut-off line) are considered unimpor-
tant.

Predictor variables sorted by importance scores

Importance Scores
Scaled Unscaled Rank Variable
100.0 1.23414E+01 1.00 age
76.0 9.38189E+00 2.00 chf
47.8 5.89639E+00 3.00 year
38.6 4.76455E+00 4.00 bmi
35.7 4.41122E+00 5.00 hr
23.7 2.92057E+00 6.00 diasbp

Wei-Yin Loh 225 GUIDE manual

6.2 Regression with censoring: heart attack data 6 IMPORTANCE SCORING

18.4 2.26483E+00 7.00 mitype

13.8 1.70883E+00 8.00 miord
12.9 1.59334E+00 9.00 sho
12.5 1.54693E+00 10.00 gender
12.2 1.49968E+00 11.00 afb
11.2 1.37849E+00 12.00 los
9.6 1.18834E+00 13.00 sysbp
----------------- cut-off -------------------
6.3 7.82079E-01 14.00 cvd
2.2 2.67578E-01 15.00 av3
Variables with unscaled scores above 1 are important

Number of important and unimportant split variables: 13, 2

The scores are also contained in the file whas500imp.scr for input into another
computer program:
Rank Score Variable
1.00 1.23414E+01 age
2.00 9.38189E+00 chf
3.00 5.89639E+00 year
4.00 4.76455E+00 bmi
5.00 4.41122E+00 hr
6.00 2.92057E+00 diasbp
7.00 2.26483E+00 mitype
8.00 1.70883E+00 miord
9.00 1.59334E+00 sho
10.00 1.54693E+00 gender
11.00 1.49968E+00 afb
12.00 1.37849E+00 los
13.00 1.18834E+00 sysbp
14.00 7.82079E-01 cvd
15.00 2.67578E-01 av3

Finally, here are the contents of the file whas500new.dsc. It puts an x against
the variables (cvd and av3 here) that are not important.
"whas500.csv"
"NA"
1
1 id x
2 age n
3 gender c
4 hr n
5 sysbp n

Wei-Yin Loh 226 GUIDE manual

 7 DIFFERENTIAL ITEM FUNCTIONING: GDS DATA

6 diasbp n
7 bmi n
8 cvd x
9 afb c
10 sho c
11 chf c
12 av3 x
13 miord c
14 mitype c
15 year c
16 admitdate x
17 disdate x
18 fdate x
19 los n
20 dstat x
21 lenfol t
22 fstat d

7 Differential item functioning: GDS data

GUIDE has an experimental option to identify important predictor variables and
items with differential item functioning (DIF) in a data set with two or more item
(dependent variable) scores. We illustrate it with a data set from Broekman et al.
(2011, 2008) and Marc et al. (2008). It consists of responses from 1978 subjects on
15 items. There are 3 predictor variables (age, education, and gender). The data
and description files are GDS.dat and GDS.dsc. Although the item responses in this
example are 0-1, GUIDE allows them to be in any ordinal (e.g., Likert) scale. The
contents of GDS.dsc are:
GDS.dat
NA
1
1 rid x
2 satis d
3 drop d
4 empty d
5 bored d
6 spirit d
7 afraid d
8 happy d
9 help d
10 home d
11 memory d
12 alive d

Wei-Yin Loh 227 GUIDE manual

 7 DIFFERENTIAL ITEM FUNCTIONING: GDS DATA

13 worth d
14 energy d
15 hope d
16 better d
17 total x
18 gender c
19 education n
20 age n
21 dxcurren x
22 sumscore x

Here is the session log to create an input file for identifying DIF items and the
important predictor variables:
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: dif.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1): 2
Name of batch output file: dif.out
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1): 2
Choose type of regression model:
1=linear, 2=quantile, 3=Poisson, 4=proportional hazards,
5=multiresponse or itemresponse, 6=longitudinal data (with T variables).
Input choice ([1:6], <cr>=1): 5
Choose option 5 for item response data.
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: GDS.dsc
Reading data description file ...
Training sample file: GDS.dat
Missing value code: NA
Records in data file start on line 1
Warning: N variables changed to S
Number of D variables = 15
D variables are:
satis
drop
empty
bored
spirit

Wei-Yin Loh 228 GUIDE manual

 7 DIFFERENTIAL ITEM FUNCTIONING: GDS DATA

afraid
happy
help
home
memory
alive
worth
energy
hope
better
Multivariate or univariate split variable selection:
Choose multivariate if there is an order among the D variables; otherwise choose univariate
Input 1 for multivariate, 2 for univariate ([1:2], <cr>=1): 2
Input 1 to normalize D variables, 2 for no normalization ([1:2], <cr>=1): 2
Input 1 for equal, 2 for unequal weighting of D variables ([1:2], <cr>=1):
Reading data file ...
Number of records in data file: 1978
Length of longest data entry: 4
Checking for missing values ...
Total number of cases: 1978
Col. no. Categorical variable #levels #missing values
18 gender 2 0
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values
Data checks complete
Creating missing value indicators
Some D variables have missing values
You can use all the data or only those with complete D values
Using only cases with complete D values will reduce the sample size
but allows the option of using PCA for split selection
Input 1 to use all obs, 2 to use obs with complete D values ([1:2], <cr>=2):
Rereading data
PCA can be used for variable selection
Do not use PCA if differential item functioning (DIF) scores are wanted
Input 1 to use PCA, 2 otherwise ([1:2], <cr>=2):
Choose option 2 because DIF scoring is desired.
#cases w/ miss. D = number of cases with all D values missing
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
1978 0 1 3 0 0 3 0 1
No. cases used for training: 1977
No. cases excluded due to 0 weight or missing D: 1
Finished reading data file
Warning: interaction tests skipped
Input expected fraction of noise variables erroneously selected ([0.00:0.99], <cr>=0.01):

Wei-Yin Loh 229 GUIDE manual

 8 TREE ENSEMBLES

Input 1 to save p-value matrix for differential item functioning (DIF), 2 otherwise ([1:2], <cr>=1)
Input file name to store DIF p-values: dif.pv
This file is useful for finding the items with DIF.
You can create a description file with the selected variables included or excluded
Input 2 to create such a file, 1 otherwise ([1:2], <cr>=1):
You can also output the importance scores and variable names to a file
Input 1 to create such a file, 2 otherwise ([1:2], <cr>=1):
Input file name: dif.scr
Input file is created!
Run GUIDE with the command: guide < dif.in

The importance scores are in the file dif.scr. They show that age is most
important and gender is least important.
Rank Score Variable
1.00 4.59054E+00 age
2.00 3.43418E+00 gender
3.00 2.36410E+00 education

The last column of dif.pv below shows that items #4 and #10 (bored and
memory) has DIF.
Item Itemname education age gender DIF
1 satis 0.747E-01 0.359E-01 0.918E-01 no
2 drop 0.157E-01 0.202E+00 0.904E+00 no
3 empty 0.547E-03 0.373E-01 0.241E+00 no
4 bored 0.563E-07 0.319E+00 0.361E+00 yes
5 spirit 0.978E+00 0.827E+00 0.261E-01 no
6 afraid 0.479E-01 0.157E-02 0.280E-02 no
7 happy 0.838E+00 0.591E+00 0.330E-01 no
8 help 0.160E-01 0.849E+00 0.384E-02 no
9 home 0.238E+00 0.181E+00 0.155E-03 no
10 memory 0.486E+00 0.000E+00 0.614E-02 yes
11 alive 0.276E+00 0.243E+00 0.416E+00 no
12 worth 0.126E+00 0.931E+00 0.650E+00 no
13 energy 0.471E+00 0.765E+00 0.203E-04 no
14 hope 0.490E+00 0.620E+00 0.224E+00 no
15 better 0.432E+00 0.476E+00 0.439E+00 no

8 Tree ensembles
A tree ensemble is a collection of trees. GUIDE has two methods of constructing an
ensemble. The preferred one is called GUIDE forest. Similar to Random Forest

Wei-Yin Loh 230 GUIDE manual

8.1 GUIDE forest: hepatitis data 8 TREE ENSEMBLES

(Breiman, 2001), it fits unpruned trees to bootstrap samples and randomly selects a
small subset of variables to search for splits at each node. There are, however, two
important differences:

1. GUIDE forest uses the unbiased GUIDE method for split selection and Random
Forest uses the biased CART method. As a result, GUIDE forest is very much
faster than Random Forest if the dependent variable is a class variable having
more than two distinct values and there are categorical predictor variables with
large numbers of categories. In addition, GUIDE forest is applicable to data
with missing values

2. Random Forest (Liaw and Wiener, 2002) requires apriori imputation of missing
values in the predictor variables but GUIDE forest does not need imputation.

A second GUIDE ensemble option is called bagged GUIDE. It fits pruned

GUIDE trees to bootstrap samples of the training data (Breiman, 1996). There
is some empirical evidence that, if there are many variables of which only a few are
useful for prediction, bagged GUIDE is slightly more accurate than GUIDE forest
(Loh, 2009, 2012). But GUIDE forest is much faster.

8.1 GUIDE forest: hepatitis data

Recall that in Section 4.4, the hepatitis data gave a null pruned tree due to 80% of
the observations belonging to one class.

8.2 Input file creation

0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: hepforest.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: hepforest.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1): 2
Input 1 for bagging, 2 for rforest: ([1:2], <cr>=2):
Input 1 for random splits of missing values, 2 for nonrandom: ([1:2], <cr>=2):
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Wei-Yin Loh 231 GUIDE manual

8.2 Input file creation 8 TREE ENSEMBLES

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: hepdsc.txt
Reading data description file ...
Training sample file: hepdat.txt
Missing value code: ?
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is CLASS
Reading data file ...
Number of records in data file: 155
Length of longest data entry: 6
Checking for missing values ...
Total number of cases: 155
Missing values found among categorical variables
Separate categories will be created for missing categorical variables
Number of classes: 2
Col. no. Categorical variable #levels #missing values
3 SEX 2 0
4 STEROID 2 1
5 ANTIVIRALS 2 0
6 FATIGUE 2 1
7 MALAISE 2 1
8 ANOREXIA 2 1
9 BIGLIVER 2 10
10 FIRMLIVER 2 11
11 SPLEEN 2 5
12 SPIDERS 2 5
13 ASCITES 2 5
14 VARICES 2 5
20 HISTOLOGY 2 0
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values
Data checks complete
Creating missing value indicators
Rereading data
Class #Cases Proportion
die 32 0.20645161
live 123 0.79354839
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
155 0 72 0 0 0 6 0 13
No. cases used for training: 155
No. cases excluded due to 0 weight or missing D: 0
Finished reading data file

Wei-Yin Loh 232 GUIDE manual

8.3 Results 8 TREE ENSEMBLES

Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3 ([1:3], <cr>=1):
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1):
Input name of file to store predicted class and probability: hepforest.fit
Input file is created!
Run GUIDE with the command: guide < hepforest.in

8.3 Results
Warning: Owing to the intrinsic randomness in forests, your results may differ from
those shown below.
Random forest of classification trees
No pruning
Data description file: hepdsc.txt
Training sample file: hepdat.txt
Missing value code: ?
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is CLASS
Number of records in data file: 155
Length of longest data entry: 6
Missing values found among categorical variables
Separate categories will be created for missing categorical variables
Number of classes: 2
Class proportions of dependent variable CLASS:
Class #Cases Proportion
die 32 0.20645161
live 123 0.79354839

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 CLASS d 2
2 AGE s 7.0000E+00 7.8000E+01
3 SEX c 2
4 STEROID c 2 1
5 ANTIVIRALS c 2
6 FATIGUE c 2 1
7 MALAISE c 2 1
8 ANOREXIA c 2 1
9 BIGLIVER c 2 10

Wei-Yin Loh 233 GUIDE manual

8.3 Results 8 TREE ENSEMBLES

10 FIRMLIVER c 2 11
11 SPLEEN c 2 5
12 SPIDERS c 2 5
13 ASCITES c 2 5
14 VARICES c 2 5
15 BILIRUBIN s 3.0000E-01 8.0000E+00 6
16 ALKPHOSPHATE s 2.6000E+01 2.9500E+02 29
17 SGOT s 1.4000E+01 6.4800E+02 4
18 ALBUMIN s 2.1000E+00 6.4000E+00 16
19 PROTIME s 0.0000E+00 1.0000E+02 67
20 HISTOLOGY c 2

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
155 0 72 0 0 0 6 0 13
No. cases used for training: 155
No. cases excluded due to 0 weight or missing D: 0

Univariate split highest priority

No interaction and linear splits
Number of trees in ensemble: 500
Number of variables used for splitting: 7
Simple node models
Estimated priors
Unit misclassification costs
Fraction of cases used for splitting each node: 0.64516
Max. number of split levels: 10
Min. node sample size: 5
Mean number of terminal nodes: 10.42

Classification matrix for training sample:

Predicted True class
class die live
die 14 10
live 18 113
Total 32 123

Number of cases used for tree construction: 155

Number misclassified: 28
Resubstitution est. of mean misclassification cost: 0.1806

Predicted class probabilities are stored in hepforest.fit

Except for the number of observations misclassified, the above results are not
particularly useful; they mostly provide a record of the parameter values chosen to
construct the forest. The predicted class probabilities in the file hepforest.fit are

Wei-Yin Loh 234 GUIDE manual

8.4 Bagged GUIDE 8 TREE ENSEMBLES

more useful, the top few lines of which are shown below. The first column indicates
whether or not the observation is used for training (labeled y vs. n), followed
by its predicted class probabilities. The last two columns give the predicted and
observed class labels. For example, observation 7 below has predicted proabilities
of 0.2702 and 0.7298 for being in class die and live, respectively, and its predicted
class is live.
train "die" "live" predicted observed
y 0.11211E-01 0.98879E+00 "live" "live"
y 0.63869E-01 0.93613E+00 "live" "live"
y 0.22287E-01 0.97771E+00 "live" "live"
y 0.84751E-02 0.99152E+00 "live" "live"
y 0.68997E-02 0.99310E+00 "live" "live"
y 0.81449E-02 0.99186E+00 "live" "live"
y 0.27020E+00 0.72980E+00 "live" "die"

8.4 Bagged GUIDE

We now apply bagged GUIDE to the same data.
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 1
Name of batch input file: hepbag.in
Input 1 for model fitting, 2 for importance or DIF scoring,
3 for data conversion ([1:3], <cr>=1):
Name of batch output file: hepbag.out
Input 1 for single tree, 2 for ensemble ([1:2], <cr>=1): 2
Input 1 for bagging, 2 for rforest: ([1:2], <cr>=2): 1
Input 1 for classification, 2 for regression, 3 for propensity score grouping
(propensity score grouping is an experimental option)
Input your choice ([1:3], <cr>=1):
Input 1 for default options, 2 otherwise ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: hepdsc.txt
Reading data description file ...
Training sample file: hepdat.txt
Missing value code: ?
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is CLASS
Reading data file ...
Number of records in data file: 155

Wei-Yin Loh 235 GUIDE manual

8.4 Bagged GUIDE 8 TREE ENSEMBLES

Length of longest data entry: 6

Checking for missing values ...
Total number of cases: 155
Missing values found among categorical variables
Separate categories will be created for missing categorical variables
Number of classes: 2
Col. no. Categorical variable #levels #missing values
3 SEX 2 0
4 STEROID 2 1
5 ANTIVIRALS 2 0
6 FATIGUE 2 1
7 MALAISE 2 1
8 ANOREXIA 2 1
9 BIGLIVER 2 10
10 FIRMLIVER 2 11
11 SPLEEN 2 5
12 SPIDERS 2 5
13 ASCITES 2 5
14 VARICES 2 5
20 HISTOLOGY 2 0
Re-checking data ...
Allocating missing value information
Assigning codes to categorical and missing values
Data checks complete
Creating missing value indicators
Rereading data
Class #Cases Proportion
die 32 0.20645161
live 123 0.79354839
Total #cases w/ #missing
#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
155 0 72 0 0 0 6 0 13
No. cases used for training: 155
No. cases excluded due to 0 weight or missing D: 0
Finished reading data file
Choose 1 for estimated priors, 2 for equal priors, 3 for priors from a file
Input 1, 2, or 3 ([1:3], <cr>=1):
Choose 1 for unit misclassification costs, 2 to input costs from a file
Input 1 or 2 ([1:2], <cr>=1):
Input name of file to store predicted class and probability: hepbag.fit
Input file is created!
Run GUIDE with the command: guide < hepbag.in

Results

Wei-Yin Loh 236 GUIDE manual

8.4 Bagged GUIDE 8 TREE ENSEMBLES

Ensemble of bagged classification trees

Pruning by cross-validation
Data description file: hepdsc.txt
Training sample file: hepdat.txt
Missing value code: ?
Records in data file start on line 1
Warning: N variables changed to S
Dependent variable is CLASS
Number of records in data file: 155
Length of longest data entry: 6
Missing values found among categorical variables
Separate categories will be created for missing categorical variables
Number of classes: 2
Class #Cases Proportion
die 32 0.20645161
live 123 0.79354839

Summary information (without x variables)

d=dependent, b=split and fit cat variable using 0-1 dummies,
c=split-only categorical, n=split and fit numerical, f=fit-only numerical,
s=split-only numerical, w=weight
Column Name Minimum Maximum #Categories #Missing
1 CLASS d 2
2 AGE s 7.0000E+00 7.8000E+01
3 SEX c 2
4 STEROID c 2 1
5 ANTIVIRALS c 2
6 FATIGUE c 2 1
7 MALAISE c 2 1
8 ANOREXIA c 2 1
9 BIGLIVER c 2 10
10 FIRMLIVER c 2 11
11 SPLEEN c 2 5
12 SPIDERS c 2 5
13 ASCITES c 2 5
14 VARICES c 2 5
15 BILIRUBIN s 3.0000E-01 8.0000E+00 6
16 ALKPHOSPHATE s 2.6000E+01 2.9500E+02 29
17 SGOT s 1.4000E+01 6.4800E+02 4
18 ALBUMIN s 2.1000E+00 6.4000E+00 16
19 PROTIME s 0.0000E+00 1.0000E+02 67
20 HISTOLOGY c 2

Total #cases w/ #missing

#cases miss. D ord. vals #X-var #N-var #F-var #S-var #B-var #C-var
155 0 72 0 0 0 6 0 13

Wei-Yin Loh 237 GUIDE manual

8.4 Bagged GUIDE 8 TREE ENSEMBLES

No. cases used for training: 155

No. cases excluded due to 0 weight or missing D: 0

Univariate split highest priority

Interaction splits 2nd priority; no linear splits
Number of trees in ensemble: 100
Pruning by v-fold cross-validation, with v = 5
Selected tree is based on mean of CV estimates
Simple node models
Estimated priors
Unit misclassification costs
Fraction of cases used for splitting each node: 0.64516
Max. number of split levels: 7
Min. node sample size: 10
Number of SEs for pruned tree: 5.0000E-01

Mean number of terminal nodes: 1.750

Classification matrix for training sample:

Predicted True class
class die live
die 0 0
live 32 123
Total 32 123

Number of cases used for tree construction: 155

Number misclassified: 32
Resubstitution est. of mean misclassification cost: 0.2065

Predicted class probabilities are stored in hepbag.fit

The top few lines of hepbag.fit follow.

train "die" "live" predicted observed
y 0.17116E+00 0.82884E+00 "live" "live"
y 0.19366E+00 0.80634E+00 "live" "live"
y 0.17912E+00 0.82088E+00 "live" "live"
y 0.17488E+00 0.82512E+00 "live" "live"
y 0.18250E+00 0.81750E+00 "live" "live"
y 0.17381E+00 0.82619E+00 "live" "live"
y 0.23391E+00 0.76609E+00 "live" "die"

Wei-Yin Loh 238 GUIDE manual

 9 OTHER FEATURES

9 Other features
9.1 Pruning with test samples
GUIDE typically has three pruning options for deciding the size of the final tree:
(i) cross-validation, (ii) test sample, and (iii) no pruning. Test-sample pruning is
available only when there are no derived variables, such as creation of dummy indi-
cator variables when b variables are present. If test-sample pruning is chosen, the
program will ask for the name of the file containing the test samples. This file must
have the same column format as the training sample file. Pruning with test-samples
or no pruning are non-default options.

9.2 Prediction of test samples

GUIDE can produce R code to predict future observations from all except kernel
and nearest neighbor classification and ensemble models. This is also a non-default
option.
Predictions of the training data for all models can be obtained, however, at the
time of tree construction. This feature can be used to obtain predictions on test
samples (i.e., observations that are not used in tree construction) by adding them
to the training sample file. There are two ways to distinguish the test observations
from the training observations:

1. Use a weight variable (designated as W in the description file) that takes value
1 for each training observation and 0 or each test observation.

2. Replace the D values of the test observations with the missing value code.

For tree construction, GUIDE does not use observations in the training sample file
that have zero weight.

9.3 GUIDE in R and in simulations

GUIDE can be used in simulations or used repeatedly on bootstrap samples to pro-
duce an ensemble of tree models. For the latter,

1. Create a file (with name data.txt, say) containing one set of bootstrapped
data.

2. Create a data description file (with name desc.txt, say) that refers to data.txt.

Wei-Yin Loh 239 GUIDE manual

9.4 Generation of powers and products 9 OTHER FEATURES

3. Create an input file (with name input.txt, say) that refers to desc.txt.

4. Write a batch program (Windows) or a shell script (Linux or Macintosh) that

repeatedly:

(a) replaces the file data.txt with new bootstrapped samples;

(b) calls GUIDE with the command: guide < input.txt; and
(c) reads and processes the results from each GUIDE run.

In R, the command in step 4b depends on the operating system. If the GUIDE

program and the files data.txt and input.txt are in the same folder as the
working R directory, the command is:

Linux/Macintosh: system("guide < input.txt > log.txt")

Windows: shell("guide < input.txt > log.txt")

If the files are not all in the same folder, full path names must be given. Here
log.txt is a text file that stores messages during execution. If GUIDE does
not run successfully, errors are also written to log.txt.

9.4 Generation of powers and products

GUIDE allows the creation of certain powers and products of regressor variables on
the fly. Specifically, variables of the form X1p X2q , where X1 and X2 are numerical
predictor variables and p and q are integers, can be created by adding one or more
lines of the form

0 i p j q a

at the end of the data description file. Here i and j are integers giving the column
numbers of variables X1 and X2 , respectively, in the data file and a is one of the
letters n, s, or f (corresponding to a numerical variable used for both splitting and
fitting, splitting only, or fitting only).
To demonstrate, suppose we wish to fit a piecewise quadratic model in the vari-
able wtgain in the birthweight data. This is easily done by adding one line to the
file birthwt.dsc. First we assign the s (for splitting only) designator to every nu-
merical predictor except wtgain. This will prevent all variables other than wtgain
from acting as regressors in the piecewise quadratic models. To create the variable
wtgain2 , add the line

Wei-Yin Loh 240 GUIDE manual

9.5 Data formatting functions 9 OTHER FEATURES

0 8 2 8 0 f

to the end of birthwt.dsc. The 8s in the above line refer to the column number of
the variables wtgain in the data file, and the f tells the program to use the variable
wtgain2 for fitting terminal node models only. Note: The line defines wtgain2 as
wtgain2 wtgain0 . Since we can equivalently define the variable by wtgain2 =
wtgain1 wtgain1 , we could also have used the line: 0 8 1 8 1 f.
The resulting description file now looks like this:
birthwt.dat
NA
1
1 weight d
2 black c
3 married c
4 boy c
5 age s
6 smoke c
7 cigsper s
8 wtgain n
9 visit c
10 ed c
11 lowbwt x
0 8 2 8 0 f

When the program is given this description file, the output will show the regression
coefficients of wtgain and wtgain2 in each terminal node of the tree.

9.5 Data formatting functions

The program includes a utility function for reformatting data files into forms required
by some statistical software packages:

1. R/Splus: Fields are space delimited. Missing values are coded as NA. Each
record is written on one line. Variable names are given on the first line.

2. SAS: Fields are space delimited. Missing values are coded with periods. Char-
acter strings are truncated to eight characters. Spaces within character strings
are replaced with underscores (_).

3. TEXT: Fields are comma delimited. Empty fields denote missing values. Char-
acter strings longer than eight characters are truncated. Each record is written
on one line. Variable names are given on the first line.

Wei-Yin Loh 241 GUIDE manual

9.5 Data formatting functions 9 OTHER FEATURES

4. STATISTICA: Fields are comma delimited. Commas in character strings are

stripped. Empty fields denote missing values. Each record occupies one line.

5. SYSTAT: Fields are comma delimited. Strings are truncated to eight char-
acters. Missing character values are replaced with spaces, missing numerical
values with periods. Each record occupies one line.

6. BMDP: Fields are space delimited. Categorical values are sorted in alphabetic
order and then assigned integer codes. Missing values are indicated by asterisks.
Variable names longer than eight characters are truncated.

7. DataDesk: Fields are space delimited. Missing categorical values are coded
with question marks. Missing numerical values are coded with asterisks. Each
record is written on one line. Spaces within categorical values are replaced with
underscores. Variable names are given on the first line of the file.

8. MINITAB: Fields are space delimited. Categorical values are sorted in alpha-
betic order and then assigned integer codes. Missing values are coded with
asterisks. Variable names longer than eight characters are truncated.

9. NUMBERS: Same as TEXT option except that categorical values are con-
verted to integer codes.

10. C4.5: This is the format required by the C4.5 (Quinlan, 1993) program.

11. ARFF: This is the format required by the WEKA (Witten and Frank, 2000)
programs.
Following is a sample session where the hepatitis data are reformatted for R or
Splus.
0. Read the warranty disclaimer
1. Create an input file for model fitting or importance scoring (recommended)
2. Convert data to other formats without creating input file
Input your choice: 3
Input name of log file: log.txt

Input 1 if D variable is categorical, 2 if real ([1:2], <cr>=1):

Input name of data description file (max 100 characters);

enclose with matching quotes if it has spaces: hepdsc.txt
Reading data description file ...
Training sample file: hepdat.txt
Missing value code: ?

Wei-Yin Loh 242 GUIDE manual

9.5 Data formatting functions 9 OTHER FEATURES

Records in data file start on line 1

Warning: N variables changed to S
Dependent variable is CLASS
Reading data file ...
Number of records in data file: 155
Length of longest data entry: 6
Checking for missing values ...
Total number of cases: 155
Number of classes = 2
Col. no. Categorical variable #levels #missing values
3 SEX 2 0
4 STEROID 2 1
5 ANTIVIRALS 2 0
6 FATIGUE 2 1
7 MALAISE 2 1
8 ANOREXIA 2 1
9 BIGLIVER 2 10
10 FIRMLIVER 2 11
11 SPLEEN 2 5
12 SPIDERS 2 5
13 ASCITES 2 5
14 VARICES 2 5
20 HISTOLOGY 2 0
Choose one of the following data formats:
Field Miss.val.codes
No. Name Separ char. numer. Remarks
------------------------------------------------------------------
1 R/Splus space NA NA 1 line/case, var names on 1st line
2 SAS space . . strings trunc., spaces -> _
3 TEXT comma empty empty 1 line/case, var names on 1st line
4 STATISTICA comma empty empty 1 line/case, commas stripped
var names on 1st line
5 SYSTAT comma space . 1 line/case, var names on 1st line
strings trunc. to 8 chars
6 BMDP space * strings trunc. to 8 chars
cat values -> integers (alph. order)
7 DATADESK space ? * 1 line/case, var names on 1st line
spaces -> _
8 MINITAB space * cat values -> integers (alph. order)
var names trunc. to 8 chars
9 NUMBERS comma NA NA 1 line/case, var names on 1st line
cat values -> integers (alph. order)
10 C4.5 comma ? ? 1 line/case, dependent variable last
11 ARFF comma ? ? 1 line/case
------------------------------------------------------------------
0 abort this job

Wei-Yin Loh 243 GUIDE manual

REFERENCES REFERENCES

Input your choice ([0:11], <cr>=1):

Input name of new data file: hep.rdata
Follow the commented lines in "hep.rdata" to read the data into R or Splus

References
Breiman, L. (1996). Bagging predictors. Machine Learning, 24:123140.

Breiman, L. (2001). Random forests. Machine Learning, 45:532.

Breiman, L., Friedman, J. H., Olshen, R. A., and Stone, C. J. (1984). Classification
and Regression Trees. Wadsworth, Belmont.

Broekman, B. F. P., Niti, M., Nyunt, M. S. Z., Ko, S. M., Kumar, R., and Ng, T. P.
(2011). Validation of a brief seven-item response bias-free Geriatric Depression
Scale. American Journal of Geriatric Psychiatry, 19:589596.

Broekman, B. F. P., Nyunt, S. Z., Niti, M., Jin, A. Z., Ko, S. M., Kumar, R.,
Fones, C. S. L., and Ng, T. P. (2008). Differential item functioning of the Geriatic
Depression Scale in an Asian population. Journal of Affective Disorders, 108:285
290.

Chan, K.-Y. and Loh, W.-Y. (2004). LOTUS: An algorithm for building accurate and
comprehensible logistic regression trees. Journal of Computational and Graphical
Statistics, 13:826852.

Chaudhuri, P. and Loh, W.-Y. (2002). Nonparametric estimation of con-

ditional quantiles using quantile regression trees. Bernoulli, 8:561576.
www.stat.wisc.edu/~ loh/treeprogs/guide/quantile.pdf.

Choi, Y., Ahn, H., and Chen, J. J. (2005). Regression trees for analysis of count
data with extra poisson variation. Computational Statistics & Data Analysis,
49(3):893915.

Hosmer, D. W., Lemeshow, S., and May, S. (2008). Applied Survival Analysis. Wiley,
2nd edition.

Hothorn, T. (2017). TH.data: THs Data Archive. R package version 1.0-8.

Ilter, N. and Guvenir, H. A. (1998). UCI machine learning repository.

Wei-Yin Loh 244 GUIDE manual

REFERENCES REFERENCES

Italiano, A. (2011). Prognostic or predictive? Its time to get back to definitions!

Journal of Clinical Oncology, 29:4718.

Kim, H. and Loh, W.-Y. (2001). Classification trees with unbiased mul-
tiway splits. Journal of the American Statistical Association, 96:589604.
www.stat.wisc.edu/~ loh/treeprogs/cruise/cruise.pdf.

Kim, H. and Loh, W.-Y. (2003). Classification trees with bivariate linear discrimi-
nant node models. Journal of Computational and Graphical Statistics, 12:512530.
www.stat.wisc.edu/~ loh/treeprogs/cruise/jcgs.pdf.

Kim, H., Loh, W.-Y., Shih, Y.-S., and Chaudhuri, P. (2007). Visualizable and
interpretable regression models with good prediction power. IIE Transactions,
39:565579. www.stat.wisc.edu/~ loh/treeprogs/guide/iie.pdf.

Koenker, R. (2005). Quantile Regression. Cambridge University Press.

Koenker, R. W. and Bassett, G. (1978). Regression quantiles. Econometrica, 46:33

50.

Koenker, R. W. and Hallock, K. (2001). Quantile regression: an introduction. Jour-

nal of Economic Perspectives, 15:143156.

Liaw, A. and Wiener, M. (2002). Classification and regression by randomforest. R

News, 2(3):1822.

Loh, W.-Y. (2002). Regression trees with unbiased variable se-

lection and interaction detection. Statistica Sinica, 12:361386.
www3.stat.sinica.edu.tw/statistica/j12n2/j12n21/j12n21.htm.

Loh, W.-Y. (2006a). Logistic regression tree analysis. In Pham, H., editor, Handbook
of Engineering Statistics, pages 537549. Springer.

Loh, W.-Y. (2006b). Regression tree models for designed experiments. In Rojo, J.,
editor, The Second Erich L. Lehmann SymposiumOptimality, volume 49, pages
210228. Institute of Mathematical Statistics Lecture Notes-Monograph Series.
arxiv.org/abs/math.ST/0611192.

Loh, W.-Y. (2008a). Classification and regression tree methods. In Rug-

geri, F., Kenett, R., and Faltin, F. W., editors, Encyclopedia of Statis-
tics in Quality and Reliability, pages 315323. Wiley, Chichester, UK.
www.stat.wisc.edu/~ loh/treeprogs/guide/eqr.pdf.

Wei-Yin Loh 245 GUIDE manual

REFERENCES REFERENCES

Loh, W.-Y. (2008b). Regression by parts: Fitting visually interpretable

models with GUIDE. In Chen, C., Hardle, W., and Unwin, A.,
editors, Handbook of Computational Statistics, pages 447469. Springer.
www.stat.wisc.edu/~ loh/treeprogs/guide/handbk.pdf.

Loh, W.-Y. (2009). Improving the precision of classifi-

cation trees. Annals of Applied Statistics, 3:17101737.
www.stat.wisc.edu/~ loh/treeprogs/guide/aoas260.pdf.

Loh, W.-Y. (2011). Classification and regression trees. Wiley Interdisciplinary Re-
views: Data Mining and Knowledge Discovery, 1:1423.

Loh, W.-Y. (2012). Variable selection for classification and regression in large p,
small n problems. In Barbour, A., Chan, H. P., and Siegmund, D., editors, Prob-
ability Approximations and Beyond, volume 205 of Lecture Notes in Statistics
Proceedings, pages 133157, New York. Springer.

Loh, W.-Y. (2014). Fifty years of classification and regression trees (with discussion).
International Statistical Review, 34:329370.

Loh, W.-Y., Chen, C.-W., and Zheng, W. (2007). Extrapolation er-

rors in linear model trees. ACM Trans. Knowl. Discov. Data, 1(2):6.
www.stat.wisc.edu/~ loh/treeprogs/guide/acm.pdf.

Loh, W.-Y., Fu, H., Man, M., Champion, V., and Yu, M. (2016). Identification of
subgroups with differential treatment effects for longitudinal and multiresponse
variables. Statistics in Medicine, 35:48374855.

Loh, W.-Y., He, X., and Man, M. (2015). A regression tree approach to identifying
subgroups with differential treatment effects. Statistics in Medicine, 34:18181833.

Loh, W.-Y. and Shih, Y.-S. (1997). Split selection meth-

ods for classification trees. Statistica Sinica, 7:815840.
www3.stat.sinica.edu.tw/statistica/j7n4/j7n41/j7n41.htm.

Loh, W.-Y. and Zheng, W. (2013). Regression trees for longitudinal and multire-
sponse data. Annals of Applied Statistics, 7:495522.

Marc, L. G., Raue, P. J., and Bruce, M. L. (2008). Screening performance of the 15-
item Geriatric Depression Scale in a diverse elderly home care population. Amer-
ican Journal of Geriatric Psychiatry, 16:914921.

Wei-Yin Loh 246 GUIDE manual

REFERENCES REFERENCES

Murnane, R. J., Boudett, K. P., and Willett, J. B. (1999). Do male dropouts ben-
efit from obtaining a GED, postsecondary education, and training? Evaluation
Reviews, 23:475502.

Peters, A. and Hothorn, T. (2015). ipred: Improved Predictors. R package version

0.9-5.

Quinlan, J. R. (1992). Learning with continuous classes. In 5th Australian Joint

Conference on Artificial Intelligence, pages 343348.

Quinlan, J. R. (1993). C4.5: Programs for Machine Learning. Morgan Kaufmann.

Rousseeuw, P. J. and Leroy, A. M. (1987). Robust Regression and Outlier Detection.

Wiley.

Schmoor, C., Olschewski, M., and Schumacher, M. (1996). Randomized and non-
randomized patients in clinical trials: experiences with comprehensive cohort stud-
ies. Statistics in Medicine, 15:263271.

Segal, M. R. (1988). Regression trees for censored data. Biometrics, 44:3547.

Singer, J. D. and Willett, J. B. (2003). Applied Longitudinal Data Analysis. Oxford

University Press, New York, NY.

Therneau, T., Atkinson, B., and Ripley, B. (2017). rpart: Recursive Partitioning
and Regression Trees. R package version 4.1-11.

Witten, I. and Frank, E. (2000). Data Mining: Practical Machine Learning Tools
and Techniques with JAVA Implementations. Morgan Kaufmann, San Fransico,
CA. www.cs.waikato.ac.nz/ml/weka.

Wei-Yin Loh 247 GUIDE manual